Using Claude Code to help me write

Using LLMs to help you write is understandably a touchy subject for many. There is quite a bit of AI slop coming out now, as it is really easy to just have the LLM tools think for you and write superficially OK but ultimately garbage prose.

My recent book, LLMs for Mortals, I used Sonnet 4.1 to write the initial draft of the book (for around $5). My prior book took around a year, whereas I was able to finish this book in around two months. I definitely did a ton of copy-editing (maybe around 20-30 hours per chapter on average), but I believe around 50% of the book material is the original Sonnet generated prose.

LLMs are a tool – they can be used poorly, but I think they can be used quite well. Pangram, a tool used to detect AI writing, does not flag any of the passages in LLMs for Mortals as AI generated.

This blog post goes over my notes on how I used Claude Code to help me write (although it really is applicable to any of the current coding tools, like Codex or Gemini as well). As a meta-reference, this blog post is 100% written by myself directly, but I will link to a draft written using Claude Code later in the post for a frame of reference.

Copy Editing

First, even if you do not agree with having an LLM write for you directly, there is a use case that should be relatively uncontroversial – having an LLM take a copy-edit pass on your work.

Here is an example I used this for recently, the blog post on Crime De-Coder goes over the benefits of using an API vs local LLMs. In this conversation, you can see my original draft, and the suggestions that Claude’s desktop tool (the free version) gave.

Again this is not really specific to Claude (this would have worked fine in ChatGPT as well). LLMs are good for not only spelling errors, but grammatical issues that spell check will not catch, as well as just more general copy-editing advice on the content.

One point of this – to replicate my setup, you need to write in plain text. Most of the things I write are in some form of markdown (plain markdown for blog posts, and Quarto for longer reports/books/etc). This makes it much easier to use the tools, especially the command line interface (CLI) tools like Claude Code.

Writing New Content

There are two big issues currently with LLM writing:

  • it is potentially wrong
  • current LLM writing has a particular style that is itself becoming noticeable

The first bullet, you need to review what it writes. It is much easier to have it write on content you are an expert in, so it is easier to review and spot errors. (It is the same current problem with using the tools to help you write computer code – they are boons for seniors but can write a ton of slop that more neophyte coders have a hard time spotting issues.)

The second bullet, having the style mimic your own, is what I am going to discuss here. It is worth understanding at a high level how generative AI LLMs work – if you ask “answer question X” vs “here is a book, …., answer question X” the LLM will generate a different response. The first part in the former prompt, “here is a book, …” is what is referred to the context. Current models have context windows (how large of a potential input) at around 500,000 words (technically they are around 1 million tokens, one word is often multiple tokens though).

You generally do not want to fill up the context window 100%, but 500,000 words is a very large number – just including text it would be multiple books. Another common prompting technique is what is called k-shot examples. It will typically go like

example input1: ...text... expected_output: ...blah...
example input2: ...diff text... expected_output: ...blah2...
....

This is what you place in the context window, then submit your usual prompt, and have the LLM generate the content. It is giving prior examples to help guide the LLM what you expect the final output to look like. This works the same way with writing – give the LLM prior examples of your writing to help it mimic your future style.

To keep it simple, I have created an example on github to follow along. Basically just have your prior writing (in text!), and then ask Claude Code something like:

review my prior blog posts in folder /blogposts, I am going to have you write a new blog post on topic X given the outline *after* you review the text

Then after your prior work is in the context window, feed the LLM an outline for what you want to write. In this example, I put the outline in an actual text file and said:

In the ClaudeWritingPost folder, review the outline.txt, then create a new md 
file, called ClaudeWritingExample.md, filling in the sections based on the 
outline

Claude Code will then go and review the text file with the outline and write the post. In the github repo I have my original outline for this same post, so you can see side-by-side.

You can technically write custom commands and skills with Claude Code (or the other CLI tools) to save the steps of typing two prompts, but to keep it simple for folks I am just showing the two steps manually. It is really just those two steps – get your prior examples into the context window, and then feed an outline for what else to write.

In the Github repo you can see some additional Claude.md files – these are files that include additional instructions. A common one I say is “do not include emojis”. LLM writing also tends to be verbose and have excessive lists. So I have instructions to avoid those as well.

The written blog post is not bad – I would suggest to go and read it as a proof of concept (I exported the session, can see it cost around fifty cents). Part of the reason I do not typically worry about blog posts is that I often add in things/change things in the process of writing. So you can see my personally written post is longer and has a few more elements.

So when would you use it? Technical writing, like writing tutorials in python, it works very well. Hence I could have it write the first pass on my LLM book and keep 50% of the content. I may use it for blog posts in the future (if I felt compelled to write something every day). But will not take that plunge for now.

For longer pieces, like an entire paper or a book, I suggest to not only make a detailed outline, but to also have the LLM write it in smaller sections. This both helps with reviewing the content, as well as to keep the LLM on track if you make edits/changes as you go. (Longer conversations it is more likely to degrade and make repeated errors.)

An Extra Note About Citations

I am not writing academic papers much anymore, but another fundamental problem with LLM writing is hallucinating citations. If you write in text markdown files, my suggestion is fairly simple – have the papers you want to cite in a bibtex file, and in-line in markdown, only cite papers in the form:

Citation, @item1 says blah [@item1; @item2]. For a specific page quote [@item1 p. 34-35].

The way I write my outlines, it typically is like write a paragraph about X, cite papers a,b,c. So my personal style of progressively filling in an outline works well with LLMs.

So this presumes you already have a list of papers (and are not using the LLM to dynamically write your lit review based on papers you have not read). Next time I actually need to write an academic paper, I may write up an MCP tool to query Semantic Scholar’s API and create a nice bibtex file.

But the solution here is again you need to review the output for accuracy. People without these tools are lazy and cite things they have not read already, so that will continue to happen (the tools just make it easier). Those that figure out how to use the tools appropriately though can be much more productive writing.

Leave a comment
by Andy Wheeler on March 20, 2026  •  Permalink
Posted in data science, Personal Productivity, writing
Tagged ,

Posted by Andy Wheeler on March 20, 2026

https://andrewpwheeler.com/2026/03/20/using-claude-code-to-help-me-write/

Some notes on the unreliability of LLM APIs

Because my book, LLMs for Mortals, was created with Quarto, it runs the code when I compile the book. It uses cached versions when no code changes, but it is guaranteed to be working code for the parts that have a grey input and a following green output, it is valid code that executed and generated the results.

I try to use temperature zero for most of the book, but some of the parts of the book are stochastic. Reasoning models you cannot set the temperature, so some elements of Chapter 3 introducing the models, and basically all of the section in chapter 6 on agents is stochastic. This actually gave me a better appreciation of some of the unreliability of these models, as for some instances it would fail, and others I needed to recompile because the output was poor.

The way jupyter caching works under the hood, it has a separate cache for the epub and the LaTeX document (that is used for the print version). So you technically get a slightly different book when you purchase epub vs paperback. When you have 60+ failure points per chapter (and that gets doubled when compiling to both epub and PDF), you get to glimpse a few of the warts of the API models.

These are also short snippets, so do not have error catching or more robust JSON parsing, so some of these issues I basically programmed away in production systems at work and did not even notice them. I figured my notes may be useful though in general for others trying to rely on these systems with large volume API calls.

OpenAI

All the models were generally reliable, but one of the examples of stochastic outputs in OpenAI gave me fits – I asked OpenAI to analyze a blog post on my Crime De-Coder site and get information from the post. Now this is a bit tricky, as the reasoning model needs to see the data is not available in the post directly, but in an image.

January 24th, at one point though this became totally unreliable in its output. It would often fail to download the additional image, and when it did, it was pretty inconsistent actually giving an accurate answer.

But now I can run below and this just returns fine and dandy near every time. Here is a loop I ran 5 times and it gives the correct answer (around 160 Tuesday at 4 AM).

from openai import OpenAI
import time

client = OpenAI()

prompt = """
Search <https://crimede-coder.com/blogposts/2024/Aoristic>, what is 
the maximum number of commercial burglaries in the chart and on what
day and hour? Do not use shorthand, give an actual number.

If you need to, download additional materials to answer the question.

Be concise in your output.
"""

for _ in range(5):
    # minimal reasoning with responses API
    response = client.responses.create(
        model="gpt-5.2",
        reasoning={'effort': 'low'},
        tools=[{"type": "web_search"}],
        input=prompt,
    )
    time.sleep(20) # to prevent going over my limit
    print(response.output_text)
    print('-------------')

My only guess is there was some downgrade in the model capabilities, and it routed behind the scenes for the reasoning models to some less capable model. (Just on January 24th though!)

Otherwise, the stochastic examples in the book using OpenAI were pretty reliable.

Anthropic

In the structured outputs chapter, I go through examples of parsing JSON vs progressively building on Pydantic outputs. I actually give examples where Pydantic schema’s can cause some filling in of data that you do not want (if the data should be null, and you use k-shot examples, it will often fill in from your last example).

So this chapter really is a ton of advice on prompt engineering for structured outputs. One example I show is using stop sequences when generating JSON and doing text parsing (which is really not necessary and best practice with Pydantic schemas, but I still use this with AWS Bedrock, since it does not support that yet).

This code works fine, what is inconsistent is that on very rare occasion, Anthropic’s API returns the bracket at the end of the call. This subsequently generates an error with this code, as it is invalid JSON with an extra bracket.

Production systems at the day job use AWS, and I wrote the text parsing in a way I would not even see this error (so not sure if it also happens with AWS). And it was quite rare with Anthropic, I just compiled the book enough times to notice this error happen on a few occasions.

Google

In the book I show off using Google Map grounding, since it is a unique capability of Googles – it was very unreliable. Not unreliable in the sense it would return an error and not be available, but unreliable in “I cannot find any google maps data right now”. So this would compile, I would just need to go look at the output and make sure it actually returned something useful.

You can see I switched to the Vertex API for this example – I cannot confidently say if Vertex was more reliable than the Gemini API for this. I experienced issues with both (maybe slightly fewer with Vertex).

The Anthropic error is not so bad – it causes an actual error in the system. The reasoning and LLM outputs something, but it is not good, troubles me more. We are really just piloting agentic systems at the day gig now with a small number of users – they have not gotten really stress tested by a large number of users. I don’t even want to think about how I would monitor maps grounding in production given my experience.

AWS

AWS I only had one example not consistently work – calling the DeepSeek API.

In the prior code calling Anthropic models via Bedrock, and later chapters I have an example of Mistral and different embedding models (Cohere and Amazon’s Titan), were all fine. Just this single example from DeepSeek would randomly not work. By not work the API would return a response, but the content would be empty. So the final print statement is where the error occurred, accessing text that did not exist.

Most of my work, even if DeepSeek is cheaper, I need to consider caching. So Haiku is pretty competitive with the other models. So I do not have much experience in Bedrock with any models besides Anthropic ones.

My biggest gripe with AWS is the IAM permissions are too difficult (and have changed over the past year). I was able to reasonably figure out how to use S3 Vectors and batch inference (which is discussed in the book). I was able to figure out Knowledge Bases, but I just took it out of the book (both too expensive for hobby projects to have the search endpoint). OpenAI’s vector search store is super easy though, so will definately consider that for traditional RAG applications moving forward.

Buy the book!

Use promo code LLMDEVS for 50% off of the epub. Or if you prefer purchase the paperback.

Leave a comment
by Andy Wheeler on February 27, 2026  •  Permalink
Posted in data science, Python
Tagged , , , ,

Posted by Andy Wheeler on February 27, 2026

https://andrewpwheeler.com/2026/02/27/some-notes-on-unreliability-of-llm-apis/

Large Language Models for Mortals book

I have published a new book, Large Language Models for Mortals: A Practical Guide for Analysts with Python. The book is available to purchase in my store, either as a paperback (for $59.99) or an epub (for $49.99).

The book is a tutorial on using python with all the major LLM foundation model providers (OpenAI, Anthropic, Google, and AWS Bedrock). The book goes through the basics of API calls, structured outputs, RAG applications, and tool-calling/MCP/agents. The book also has a chapter on LLM coding tools, with example walk throughs for GitHub Copilot, Claude Code (including how to set it up via AWS Bedrock), and Google’s Antigravity editor. (It also has a few examples of local models, which you can see Chapter 2 I discuss them before going onto the APIs in Chapter 3).

You can review the first 60 some pages (PDF link here if on Iphone).

llms_mortals_previewDownload

While many of the examples in the book are criminology focused, such as extracting out crime elements from incident narratives, or summarizing time series charts, the lessons are more general and are relevant to anyone looking to learn the LLM APIs. I say “analyst” in the title, but this is really relevant to:

  • traditional data scientists looking to expand into LLM applications
  • PhD students (in all fields) who would like to use LLM applications in their work
  • analysts looking to process large amounts of unstructured textual data

Basically anyone who wants to build or create LLM applications, this is the book to help you get started.

I wrote this book partially out of fear – the rapid pace of LLM development has really upended my work as a data scientist. It is really becoming the most important set of skills (moreso than traditional predictive machine learning) in just the past year or two. This book is the one I wish I had several years ago, and will give analysts a firm grounding in using LLMs in realistic applications.

Again, the book is available in:

For purchase worldwide. Here are all the sections in the book – whether you are an AWS or Google shop, or want to learn the different database alternatives for RAG, or want more self contained examples of agents with python code examples for OpenAI, Anthropic, or Google, this should be a resource you highly consider purchasing.

To come are several more blog posts in the near future, how I set up Claude Code to help me write (and not sound like a robot). How to use conformal inference and logprobs to set false positive rates for classification with LLM models, and some pain points with compiling a Quarto book with stochastic outputs (and points of varying reliability for each of the models).

But for now, just go and purchase the book!

Below is the table of contents to review – it is over 350 pages for the print version (in letter paper), over 250 python code snippets and over 80 screenshots.

Large Language Models for Mortals: A Practical Guide for Analysts with Python
by Andrew Wheeler
TABLE OF CONTENTS
Preface
    Are LLMs worth all the hype?
    Is this book more AI Slop?
    Who this book is for
    Why write this book?
    What this book covers
    What this book is not
    My background
    Materials for the book
    Feedback on the book
    Thank you
1 Basics of Large Language Models
    1.1 What is a language model?
    1.2 A simple language model in PyTorch
    1.3 Defining the neural network
    1.4 Training the model
    1.5 Testing the model
    1.6 Recapping what we just built
2 Running Local Models from Hugging Face
    2.1 Installing required libraries
    2.2 Downloading and using Hugging Face models
    2.3 Generating embeddings with sentence transformers
    2.4 Named entity recognition with GLiNER
    2.5 Text Generation
    2.6 Practical limitations of local models
3 Calling External APIs
    3.1 GUI applications vs API access
    3.2 Major API providers
    3.3 Calling the OpenAI API
    3.4 Controlling the Output via Temperature
    3.5 Reasoning
    3.6 Multi-turn conversations
    3.7 Understanding the internals of responses
    3.8 Embeddings
    3.9 Inputting different file types
    3.10 Different providers, same API
    3.11 Calling the Anthropic API
    3.12 Using extended thinking with Claude
    3.13 Inputting Documents and Citations
    3.14 Calling the Google Gemini API
    3.15 Long Context with Gemini
    3.16 Grounding in Google Maps
    3.17 Audio Diarization
    3.18 Video Understanding
    3.19 Calling the AWS Bedrock API
    3.20 Calculating costs
4 Structured Output Generation
    4.1 Prompt Engineering
    4.2 OpenAI with JSON parsing
    4.3 Assistant Messages and Stop Sequences
    4.4 Ensuring Schema Matching Using Pydantic
    4.5 Batch Processing For Structured Data Extraction using OpenAI
    4.6 Anthropic Batch API
    4.7 Google Gemini Batch
    4.8 AWS Bedrock Batch Inference
    4.9 Testing
    4.10 Confidence in Classification using LogProbs
    4.11 Alternative inputs and outputs using XML and YAML
    4.12 Structured Workflows with Structured Outputs
5 Retrieval-Augmented Generation (RAG)
    5.1 Understanding embeddings
    5.2 Generating Embeddings using OpenAI
    5.3 Example Calculating Cosine similarity and L2 distance
    5.4 Building a simple RAG system
    5.5 Re-ranking for improved results
    5.6 Semantic vs Keyword Search
    5.7 In-memory vector stores
    5.8 Persistent vector databases
    5.9 Chunking text from PDFs
    5.10 Semantic Chunking
    5.11 OpenAI Vector Store
    5.12 AWS S3 Vectors
    5.13 Gemini and BigQuery SQL with Vectors
    5.14 Evaluating retrieval quality
    5.15 Do you need RAG at all?
6 Tool Calling, Model Context Protocol (MCP), and Agents
    6.1 Understanding tool calling
    6.2 Tool calling with OpenAI
    6.3 Multiple tools and complex workflows
    6.4 Tool calling with Gemini
    6.5 Returning images from tools
    6.6 Using the Google Maps tool
    6.7 Tool calling with Anthropic
    6.8 Error handling and model retry
    6.9 Tool Calling with AWS Bedrock
    6.10 Introduction to Model Context Protocol (MCP)
    6.11 Connecting Claude Desktop to MCP servers
    6.12 Examples of Using the Crime Analysis Server in Claude Desktop
    6.13 What are Agents anyway?
    6.14 Using Multiple Tools with the OpenAI Agents SDK
    6.15 Composing and Sequencing Agents with the Google Agents SDK
    6.16 MCP and file searching using the Claude Agents SDK
    6.17 LLM as a Judge
7 Coding Tools and AI-Assisted Development
    7.1 Keeping it real with vibe coding
    7.2 VS Code and GitHub Install
    7.3 GitHub Copilot
    7.4 Claude Code Setup
    7.5 Configuring API access
    7.6 Using Claude Code to Edit Files
    7.7 Project context with CLAUDE.md
    7.8 Using an MCP Server
    7.9 Custom Commands and Skills
    7.10 Session Management
    7.11 Hooks for Testing
    7.12 Claude Headless Mode
    7.13 Google Antigravity
    7.14 Best practices for AI-assisted coding
8 Where to next?
    8.1 Staying current
    8.2 What to learn next?
    8.3 Forecasting the near future of foundation models
    8.4 Final thoughts
Leave a comment
by Andy Wheeler on February 11, 2026  •  Permalink
Posted in Crime Analysis, data science, Python, writing
Tagged ,

Posted by Andy Wheeler on February 11, 2026

https://andrewpwheeler.com/2026/02/11/large-language-models-for-mortals-book/

Part time product design positions to help with AI companies

Recently on the Crime Analysis sub-reddit an individual posted about working with an AI product company developing a tool for detectives or investigators.

The Mercor platform has many opportunities that may be of interest to my network, so I am sharing them here. These include not only for investigators, but GIS analysts, writers, community health workers, etc. (The eligibility interviewers I think if you had any job in gov services would likely qualify, it is just reviewing questions.)

All are part time (minimum of 15 hours per week), remote, and can be in the US, Canada, or UK. (But cannot support H1-B or OPT visas in the US).

Additional for professionals looking to get into the tech job market, see these two resources:

I actually just hired my first employee at Crime De-Coder. Always feel free to reach out if you think you would be a good fit for the types of applications I am working on (python, GIS, crime analysis experience). I will put you in the list to reach out to when new opportunities are available.

Detectives and Criminal Investigators

Referral Link

$65-$115 hourly

Mercor is recruiting Detectives and Criminal Investigators to work on a research project for one of the world’s top AI companies. This project involves using your professional experience to design questions related to your occupation as a Detective and Criminal Investigator. Applicants must:

  • Have 4+ years full-time work experience in this occupation;
  • Be based in the US, UK, or Canada
  • minimum of 15 hours per week

Community Health Workers

Referral Link

$60-$80 hourly

Mercor is recruiting Community Health Workers to work on a research project for one of the world’s top AI companies. This project involves using your professional experience to design questions related to your occupation as a Community Health Worker. Applicants must:

  • Have 4+ years full-time work experience in this occupation;
  • Be based in the US, UK, or Canada
  • minimum 15 hours per week

Writers and Authors

Referral Link

$60-$95 hourly

Mercor is recruiting Writers and Authors to work on a research project for one of the world’s top AI companies. This project involves using your professional experience to design questions related to your occupation as a Writer and Author.

Applicants must:

  • Have 4+ years full-time work experience in this occupation;
  • Be based in the US, UK, or Canada
  • minimum 15 hours per week

Eligibility Interviewers, Government Programs

Referral Link

$60-$80 hourly

Mercor is recruiting Eligibility Interviewers, Government Programs to work on a research project for one of the world’s top AI companies. This project involves using your professional experience to design questions related to your occupation as a Eligibility Interviewers, Government Program. Applicants must:

  • Have 4+ years full-time work experience in this occupation;
  • Be based in the US, UK, or Canada
  • minimum 15 hours per week

Cartographers and Photogrammetrists

Referral Link

$60-$105 hourly

Mercor is recruiting Cartographers and Photogrammetrists to work on a research project for one of the world’s top AI companies. This project involves using your professional experience to design questions related to your occupation as a Cartographer and Photogrammetrist. Applicants must:

  • Have 4+ years full-time work experience in this occupation;
  • Be based in the US, UK, or Canada
  • minimum 15 hours per week

Geoscientists, Except Hydrologists and Geographers

$85-$100 hourly

Referral Link

Mercor is recruiting Geoscientists, Except Hydrologists and Geographers to work on a research project for one of the world’s top AI companies. This project involves using your professional experience to design questions related to your occupation as a Geoscientists, Except Hydrologists and Geographers Applicants must:

  • Have 4+ years full-time work experience in this occupation;
  • Be based in the US, UK, or Canada
  • minimum of 15 hours per week
Leave a comment
by Andy Wheeler on January 15, 2026  •  Permalink
Posted in Crime Analysis, Crime Mapping, Criminal Justice, data science, writing
Tagged , ,

Posted by Andy Wheeler on January 15, 2026

https://andrewpwheeler.com/2026/01/15/part-time-product-design-positions-to-help-with-ai-companies/

Year in Review 2025 and AI Predictions

For a brief year in review, total views for the two different websites have decreased in the past year. For this blog, I am going to be a few thousand shy of 100,000 views. (2023 I had over 150k views, and 2024 I had over 140k views.) For the Crime De-Coder site, I am going to only get around 15k views.

Part of it is I posted less, this will be the 21st blog post this year on the personal blog (2023 had 46 and 2024 had 32 posts). The Crime De-Coder site had 12 blog posts, so pretty consistent with the prior year. Both are pretty bursty, with large bouts of traffic coming from if I post something to Hacker News I can get 1k to 10k views in a day or two if it makes it to the front page. So the 2024 stats for the crime de-coder was a few of those Hacker News bumps I did not get in 2025.

Some of it could legitimately be traditional Google search being usurped by the gen AI tools. This is the first year I had appreciable referrals from chatgpt, but they are less than 1000. The other tools are trivial amount of referrals. If I worried about SEO more, I would have more updating/regular content (as old pages are devalued quite a bit by google, and it seems to be getting more severe over time).

I have upped my use of the free tools quite a bit. ChatGPT knows me pretty well, and I use Claude Desktop almost every day as well.

An IAM policy scroll is more of a nightmare, and I definitely ask more python questions than R, but the cartoon desk is pretty close to spot on. I am close to paying for Anthropic subscription for Claude code credits (currently use pay as I go via Bedrock, and this is the first month I went over $20).

What pages on the blog are popular I can never be sure of. My most popular post last year was Downloading Police Employment Trends from the FBI Data Explorer. A 2023 post, that had random times where it would have several hundred visits in a short hour span. (Some bot collecting sites? I do not know.) If it is actual people, you would want to check out my Sworn Dashboard site, where you can look at trends for PDs much easier than downloading all the data yourself!

One thing that has grown though, I do short form posting on LinkedIn on my crime de-coder page. Impressions total for the year is over 340k (see the graph), and I currently am a few shy of 4400 followers.

LinkedIn is nice because it can be slightly longer form than the other social media sites. I would suggest you follow me there (in addition to signing up for RSS feeds for the two sites). That is the easiest way to follow my work.

I also took over as a moderator of the Crime Analysis Reddit forum, it is better than the IACA forums in my opinion, so encourage folks to post there for crime analysis questions.

Crime De-Coder Work

Crime De-Coder work has been steady (but not increasing). Similar to last year had several consulting gigs conducting crime analysis for premises liability cases (and one other case I may share my opinions once it is over), and doing some small projects with non-profits and police departments.

One big project was a python training in Austin.

The Python Book (which I also translated to Spanish/French), had a trickle of new sales. 2024 had around 100 sales and 2025 had around 50 sales. It is close to 2/3 print sales and 1/3 epub, so definately folks should have physical prints if you are selling books still.

Doing trainings basically makes writing the book worth it, but I do hope eventually the book makes it way into grad school curriculum’s. (Only one course so far.) I have pitched to grad schools to have me run a similar bootcamp to what I do for crime analysts, so if interested let me know.

The biggest new thing was Crime De-Coder got an Arnold Grant. Working with Denver PD on an experiment to evaluate a chronic offender initiative.

At the Day Gig

At my day gig, I was officially promoted to a senior manager and then quickly to a director position. Hence you get posts like what to show in your tech resume and notes on project management.

One of the reasons I am big on python – it is the dominant programming language in data science. It is hard for me to recruit from my network, as majority of individuals just know a little R (if you were a hard core R person, had packages/well executed public repo’s, I could more easily think you will be able to migrate to python to work on my team).

So learn python if you want to be a data scientist is my advice (and see other job market advice at my archived newsletter).

AI Predictions

At the day gig, my work went from 100% traditional supervised machine learning models to more like 50/50 traditional vs generative AI applications. The genAI hype is real, but I think it is worthwhile putting my thoughts to paper.

The biggest question is will AI take all of our jobs? I think a more likely end scenario is the AI tools just become better at helping humans do tasks. The leap from helping a human do something faster vs an AI tool doing it 100% on its own with 0 human input is hard. The models are getting incrementally better, but I think to fully replace people in a substantive way will require another big advancement in fundamental capabilities. Making a human 10x more productive is easier and still will make the AI companies a ton of money.

Sometimes people view the 10x idea and say that will take jobs, just not 100% of jobs. That is a view though that there is only a finite amount of work to be done. That assumption is clearly not true, and being able to do work faster/cheaper just induces demand for more potential work. The example with calculators making more banking jobs, not less, is basically the same example.

One of the critiques of the current systems is they are overvalued, so we are in a bubble. I do not remember where I read it, but one estimate was if everyone in the US spent $1 a day on the different AI tools, that would justify the current valuations for OpenAI, Anthropic, NVIDIA, etc. I think that is totally doable, we spend a few thousand a workday at Gainwell on the foundation models for example for a few projects, and we are just going to continue to roll out more and more. Gainwell is a company with around 6k employees for reference, and our current AI applications touch way less than 1k of those employees. We have plenty of room to grow those applications.

It is super hard though to build systems to help people do things faster. And we are talking like “this thing that used to take 30 minutes now takes 15 minutes”. If you have 100 people doing that thing all the time though, the costs of the models are low enough it is an easy win.

And this mostly only holds true for knowledge economy work that can be all done via software. There just still needs to be fundamental improvements to robotics to be able to do physical things. The tailor’s job is safe for the foreseeable future.

The change in the data science landscape to more generative AI applications definitely requires social scientists and analysts to up their game though to learn a new set of tools. I do have another book in the works to address that, so hopefully you will see that early next year.

Leave a comment
by Andy Wheeler on December 24, 2025  •  Permalink
Posted in Crime Analysis, Criminal Justice, Personal Productivity, Python, R, scholarly
Tagged

Posted by Andy Wheeler on December 24, 2025

https://andrewpwheeler.com/2025/12/24/year-in-review-2025-and-ai-predictions/

Advice for crime analyst to break into data science

I recently received a question about a crime analyst looking to break into data science. Figured it would be a good topic for my advice in a blog post. I have written many resources over the years targeting recent PhDs, but the advice for crime analysts is not all that different. You need to pick up some programming, and likely some more advanced tech skills.

For background, the individual had SQL + Excel skills (which many analysts may just have Excel). Vast majority of analyst roles, you should be quite adept at SQL. But just SQL is not sufficient for even an entry level data science role.

For entry data science, you will need to demonstrate competency in at least one programming language. The majority of positions will want you to have python skills. (I wrote an entry level python book exactly for someone in your position.)

You likely will also need to demonstrate competency in some machine learning or using large language models for data science roles. It used to be Andrew Ng’s courses were the best recommendation (I see he has a spin off DeepLearningAI now). So that is second hand though, I have not personally taken them. LLMs are more popular now, so prioritizing learning how to call those APIs, build RAG systems, prompt engineering I think is going to make you slightly more marketable than traditional machine learning.

I have personally never hired anyone in a data science role without a masters. That said, I would not have a problem if you had a good portfolio. (Nice website, Github contributions, etc.)

You should likely start just looking and applying to “analyst” roles now. Don’t worry about if they ask for programming you do not have experience in, just apply. Many roles the posting is clearly wrong or totally unrealistic expectations.

Larger companies, analyst roles can have a better career ladder, so you may just decide to stay in that role. If not, can continue additional learning opportunities to pursue a data science career.

Remote is more difficult than in person, but I would start by identifying companies that are crime analysis adjacent (Lexis Nexis, ESRI, Axon) and start applying to current open analyst positions.

For additional resources I have written over the years:

The alt-ac newsletter has various programming and job search tips. THe 2023 blog post goes through different positions (if you want, it may be easier to break into project management than data science, you have a good background to get senior analyst positions though), and the 2025 blog post goes over how to have a portfolio of work.

Cover page, data science for crime analysis with python
Leave a comment
by Andy Wheeler on November 21, 2025  •  Permalink
Posted in Crime Analysis, Crime Mapping, Criminal Justice, data science, Python
Tagged

Posted by Andy Wheeler on November 21, 2025

https://andrewpwheeler.com/2025/11/21/advice-for-crime-analyst-to-break-into-data-science/

Why I like working at Gainwell

There is an opening for a principal data science position on my team at Gainwell. I was promoted to a director role a few months ago (I was an individual contributor for 4+ years), so this position will directly report to me.

There are of course mixed reviews if you google Gainwell Technologies what it is like to work there. Gainwell is a big organization (maybe over 6,000 employees at this point). Any big org it can be totally different if you work for a different team. So I can describe what it is like to work for me and on my team.

I have been at HMS (which then merged with Gainwell) since the end of 2019. The way I describe what Gainwell does, we administer Medicaid claims for many states, but most of what my team does is support efforts to fix bad Medicaid claims (more commonly referred to as Fraud/Waste/Abuse). Fix means for the most part the bill was wrong (too high or for the wrong services), or the bill should have gone to a different person (seems trivial, but billions are sent to Medicaid that should be sent to a commercial carrier or car insurance in the US per year).

I started as a data scientist, went up through a ranks with a promotion every few years (our levels are JR->Advisor->Principal->Sr Manager->Director). For a list of projects folks on my team are working on:

  • structured labeling for medical records to help with audits
  • supervised machine learning models to help make prior authorizations go faster
  • structured extraction for many types of documents to auto-verify the contents
  • improving supervised machine learning models in production to improve identification of third parties liable for claims
  • building tools to auto-identify potential Fraud/Waste/Abuse in large databases of claims
  • making our master name index record linkage project run much faster and have more matches

This specific role will be to help with the last bullet. So it is a mix of the more recent fad of GenAI tools, but also we do a bunch of traditional machine learning as well. This role I want folks to know at least two {SQL,Machine Learning,Python}.

We are mostly just solution architects, using python to glue together different processes to make smart decisions across the org. Gainwell is a very federated and older company built up from acquisitions over the years – my team is one of the few that works across the org with many different teams. While many people at Gainwell have a data science title, my team is the AIML (artificial intelligence and machine learning) team at Gainwell.

This work is important, just my team is associated with models that help save states 9 figures on Medicaid claims per year (I imagine it is well over 10 figures across the whole org). For a bit more about the team:

  • all remote, we have a mix across the continental US. Most of Gainwell is in central, so often the schedule follows central time for meetings (think many stand-ups at 9:30-10am eastern)
  • my team has a bunch of smaller groups working on individual projects (think 2 or 3 max assigned to work on these projects). This position will be number 8 under me.
  • We are always tied to revenue when we take on a project (think generally if we cannot justify at minimum 1e6 in savings or increased revenue, we will not take on that project). This is important, as we are not just treated like IT filling tickets, we are deeply integrated with the business teams we are delivering solutions for.
  • You do not have to make sales, we are building things internally (think your clients are different teams inside of Gainwell). We have more work than we can do.
  • We are committed to building things fast. Those first 4 bullets are things we built in the past 6 months and are already generating millions of dollars of revenue.

So we are working with legacy on-premises Hadoop systems, or Databricks, or AWS deployments calling Bedrock or openai, or traditional machine learning – we just get a new project and figure it out. Do not take this as you need to know everything (if you put some random tech on your resume like Docker or time series analysis, be prepared to answer a random question about it when I interview you!) But I do want smart folks who can learn and adapt to the situation at hand. Really the only consistent tech across projects is python and SQL.

For work life balance:

  • I have worked on the weekend a total of 2 times in my tenure at Gainwell that I can remember (and those were less than 1 hour)
  • Because we are remote, we are uber flexible with scheduling (need to drop off or pick up your kid from school during the day, can very likely work around that schedule). Basically all I care about is you get your work done, and individual contributors will just need to figure out a few regular meetings (like stand ups or other meetings with the business group you are working with)
  • Flex PTO (basically only need permission if you are taking more than a week off)

I shared this on LinkedIn and I slightly dread it (I will get a million messages) – but readers of this blog are different. If you think you may be interested in private sector, feel free to reach out. (Lets just get criminologists to take over, Gio now also has his own team at Gainwell.)

Since this is a principal position, recent grads will be tougher to advocate for, but I can give feedback what I am looking for (and it is good for you to be on my radar for future positions). If you are a superstar and the salary is not enough, reach out and we can have a one on one chat.

The main thing is most of my network uses R, and not python. Basically the way I view it is a weighted scale. If you are really good at R (have public code examples and packages), I can be more confident you can learn python. If you knew python, and just had the typical replication code sets in python (a much lower bar). No excuses to learn python, I wrote a book to help with that.

Leave a comment
by Andy Wheeler on November 13, 2025  •  Permalink
Posted in Uncategorized

Posted by Andy Wheeler on November 13, 2025

https://andrewpwheeler.com/2025/11/13/why-i-like-working-at-gainwell/

What to show in your tech resume?

Jason Brinkley on LinkedIn the other day had a comment on the common look of resumes – I disagree with his point in part but it is worth a blog post to say why:

So first, when giving advice I try to be clear about what I think are just my idiosyncratic positions vs advice that I feel is likely to generalize. So when I say, you should apply to many positions, because your probability of landing a single position is small, that is quite general advice. But here, I have personal opinions about what I want to see in a resume, but I do not really know what others want to see. Resumes, when cold applying, probably have to go through at least two layers (HR/recruiter and the hiring manager), who each will need different things.

People who have different colored resumes, or in different formats (sometimes have a sidebar) I do not remember at all. I only care about the content. So what do I want to see in your resume? (I am interviewing for mostly data scientist positions.) I want to see some type of external verification you actually know how to code. Talk is cheap, it is easy to list “I know these 20 python libraries” or “I saved our company 1 million buckaroos”.

So things I personally like seeing in a resume are:

  • code on github that is not a homework assignment (it is OK if unfinished)
  • technical blog posts
  • your thesis! (or other papers you were first/solo author)

Very few people have these things, so if you do and you land in my stack, you are already at the like 95th percentile (if not higher) for resumes I review for jobs.

The reason having outside verification you actually know what you are doing is because people are liars. For our tech round, our first question is “write a python hello world program and execute it from the command line” – around half of the people we interview fail this test. These are all people who list they are experts in machine learning, large language models, years of experience in python, etc.

My resume is excessive, but I try to practice what I preach (HTML version, PDF version)

I added some color, but have had recruiters ask me to take it off the resume before. So how many people actually click all those links when I apply to positions? Probably few if any – but that is personally what I want to see.

There are really only two pieces of advice I have seen repeatedly about resumes that I think are reasonable, but it is advice not a hard rule:

  • I have had recruiters ask for specific libraries/technologies at the top of the resume
  • Many people want to hear about results for project experience, not “I used library X”

So while I dislike the glut of people listing 20 libraries, I understand it from the point of a recruiter – they have no clue, so are just trying to match the tech skills as best they can. (The matching at this stage I feel may be worse than random, in that liars are incentivized, hence my insistence on showing actual skills in some capacity.) It is infuriating when you have a recruiter not understand some idiosyncratic piece of tech is totally exchangeable with what you did, or that it is trivial to learn on the job given your prior experience, but that is not going to go away anytime soon.

I’d note at Gainwell we have no ATS or HR filtering like this (the only filtering is for geographic location and citizenship status). I actually would rather see technical blog posts or personal github code than saying “I saved the company 1 million dollars” in many circumstances, as that is just as likely to be embellished as the technical skills. Less technical hiring managers though it is probably a good idea to translate technical specs to more plain business implications though.

2 Comments
by Andy Wheeler on November 1, 2025  •  Permalink
Posted in data science, Personal Productivity, Python, scholarly
Tagged

Posted by Andy Wheeler on November 1, 2025

https://andrewpwheeler.com/2025/11/01/what-to-show-in-your-tech-resume/

I translated my book for $7 using openai

The other day an officer from the French Gendarmerie commented that they use my python for crime analysis book. I asked that individual, and he stated they all speak English. But given my book is written in plain text markdown and compiled using Quarto, it is not that difficult to pipe the text through a tool to translate it to other languages. (Knowing that epubs under the hood are just html, it would not suprise me if there is some epub reader that can use google translate.)

So you can see now I have available in the Crime De-Coder store four new books:

ebook versions are normally $39.99, and print is $49.99 (both available worldwide). For the next few weeks, can use promo code translate25 (until 11/15/2025) to purchase epub versions for $19.99.

If you want to see a preview of the books first two chapters, here are the PDFs:

And here I added a page on my crimede-coder site with testimonials.

As the title says, this in the end cost (less than) $7 to convert to French (and ditto to convert to Spanish).

Here is code demo’ing the conversion. It uses OpenAI’s GPT-5 model, but likely smaller and cheaper models would work just fine if you did not want to fork out $7. It ended up being a quite simple afternoon project (parsing the markdown ended up being the bigger pain).

So the markdown for the book in plain text looks like this:

It ends up that because markdown uses line breaks to denote different sections, that ends up being a fairly natural break to do the translation. These GenAI tools cannot repeat back very long sequences, but a paragraph is a good length. Long enough to have additional context, but short enough for the machine to not go off the rails when trying to just return the text you input. Then I just have extra logic to not parse code sections (that start/end with three backticks). I don’t even bother to parse out the other sections (like LaTeX or HTML), and I just include in the prompt to not modify those.

So I just read in the quarto document, split by “”, then feed in the text sections into OpenAI. I did not test this very much, just use the current default gpt-5 model with medium reasoning. (It is quite possible a non-reasoning smaller model will do just as well. I suspect the open models will do fine.)

You will ultimately still want someone to spot check the results, and then do some light edits. For example, here is the French version when I am talking about running code in the REPL, first in English:

Running in the REPL

Now, we are going to run an interactive python session, sometimes people call this the REPL, read-eval-print-loop. Simply type python in the command prompt and hit enter. You will then be greeted with this screen, and you will be inside of a python session.

And then in French:

Exécution dans le REPL

Maintenant, nous allons lancer une session Python interactive, que certains appellent le REPL, boucle lire-évaluer-afficher. Tapez simplement python dans l’invite de commande et appuyez sur Entrée. Vous verrez alors cet écran et vous serez dans une session Python.

So the acronym is carried forward, but the description of the acronym is not. (And I went and edited that for the versions on my website.) But look at this section in the intro talking about GIS:

There are situations when paid for tools are appropriate as well. Statistical programs like SPSS and SAS do not store their entire dataset in memory, so can be very convenient for some large data tasks. ESRI’s GIS (Geographic Information System) tools can be more convenient for specific mapping tasks (such as calculating network distances or geocoding) than many of the open source solutions. (And ESRI’s tools you can automate by using python code as well, so it is not mutually exclusive.) But that being said, I can leverage python for nearly 100% of my day to day tasks. This is especially important for public sector crime analysts, as you may not have a budget to purchase closed source programs. Python is 100% free and open source.

And here in French:

Il existe également des situations où les outils payants sont appropriés. Les logiciels statistiques comme SPSS et SAS ne stockent pas l’intégralité de leur jeu de données en mémoire, ils peuvent donc être très pratiques pour certaines tâches impliquant de grands volumes de données. Les outils SIG d’ESRI (Système d’information géographique) peuvent être plus pratiques que de nombreuses solutions open source pour des tâches cartographiques spécifiques (comme le calcul des distances sur un réseau ou le géocodage). (Et les outils d’ESRI peuvent également être automatisés à l’aide de code Python, ce qui n’est pas mutuellement exclusif.) Cela dit, je peux m’appuyer sur Python pour près de 100 % de mes tâches quotidiennes. C’est particulièrement important pour les analystes de la criminalité du secteur public, car vous n’avez peut‑être pas de budget pour acheter des logiciels propriétaires. Python est 100 % gratuit et open source.

So it translated GIS to SIG in French (Système d’information géographique). Which seems quite reasonable to me.

I paid an individual to review the Spanish translation (if any readers are interested to give me a quote for the French version copy-edits, would appreciate it). She stated it is overall very readable, but just has many minor things. Here is a a sample of suggestions:

Total number of edits she suggested were 77 (out of 310 pages).

If you are interested in another language just let me know. I am not sure about translation for the Asian languages, but I imagine it works OK out of the box for most languages that are derivative of Latin. Another benefit of self-publishing, I can just have the French version available now, but if I am able to find someone to help with the copy-edits I will just update the draft after I get their feedback.

Leave a comment
by Andy Wheeler on October 25, 2025  •  Permalink
Posted in Crime Analysis, Crime Mapping, data science, Python, writing
Tagged , , ,

Posted by Andy Wheeler on October 25, 2025

https://andrewpwheeler.com/2025/10/25/i-translated-my-book-for-7-using-openai/

Recording your mouse and keyboard with python

The differently LLM providers have released computer use tools, Google’s Gemini being one of the most recent ones. They way these work, it submits and image, and then does tasks given general instruction, actually manipulating the mouse and keyboard, asking for human in the loop input at various stages, etc. They seem cool but I am lacking clear examples where I would use them.

They really make sense for very complicated workflows that vary. The things that make the most sense to spend some time trying to automate though are boring things – things that have a specific set of inputs and outputs, and the machine can entirely make those route decisions based on rules you define at the onset. When I was a crime analyst, I worked with all sorts of janky and old desktop software that I would need to generate reports and email those results to key individuals for example.

Here as a day project, I created a set of python functions to record your keyboard and mouse inputs, and then replay those inputs. The python code is here, it ends up being pretty simple (mostly written by ChatGPT and slightly modified by me). And here is a YouTube video demonstration showing what you can do.

So even if you have software that does not have an easy way to integrate with other code, here you can just record your movements and replay them with python on a schedule to accomplish that monotonous task.

Leave a comment
by Andy Wheeler on October 10, 2025  •  Permalink
Posted in Uncategorized

Posted by Andy Wheeler on October 10, 2025

https://andrewpwheeler.com/2025/10/10/recording-your-mouse-and-keyboard-with-python/

« Older Posts