Applied AI engineering
AI engineering.
For answers you
can stand behind.
I build language-model features for products where a wrong answer costs something. Retrieval that cites its sources. Checks written in code. Evals that tell you whether a change helped.
Joel Stevick · Independent engineer & author
20 years shipping software · Frostburg, Maryland
Answers checked against the sources they cite.
Selected projects
The work, in production.
All in production. Each one is something I designed and built.
Howard's Lantern: a question box that cites the law
A driving-law exam-prep app whose ask box answers a learner's question in two sentences from their own state's manual or vehicle code, with the section underneath.
- 172,486 law rows across all 51 states, fetched verbatim, dated and hashed
- BM25 and vector search fused by rank, a Haiku rephraser and reranker, then a check that refuses any figure not in a cited law
- A 190-case eval against the live model, run as a gate on pull requests
PalabraEdu: real-time education SaaS
A five-language multiplayer learning platform on Cloudflare Workers and Durable Objects, with payments, notifications and teacher dashboards.
- Built in six weekends
- 110,000 lines, 6,200+ automated tests
Quorum
An open-source Claude Code plugin that makes a coding agent write a falsifiable plan, then runs six independent review lenses, a judge, and a CI guard before a pull request opens.
Multi-tenant RAG platform on AWS
Hybrid Pinecone retrieval, Cohere reranking, Bedrock model routing, query decomposition and streamed answers, with the whole stack in AWS CDK.
On-device document intelligence
Commercial PII-redaction and RAG authoring apps for macOS and Windows: DistilBERT on ONNX, searchable PDFs, audit trails, signed offline licensing.
Work with me
A clear scope. A working result.
Fixed prices with a written scope: what is in, what is out, and what you own at the end. Inside the scope the price does not move.
AI chatbot for your website
A chatbot that answers from your documents, shows its sources, and says when it cannot find an answer. One website widget, up to 50 text-based documents and 500,000 extracted words, a one-time import, and 30 agreed acceptance questions. Deployed in your account with source code and a handover guide.
Grounding and eval review
I read your RAG or LLM feature end to end and tell you where it can show an answer nobody checked. You get a written report, a golden set to measure against, and the fixes ranked by risk.
Build a retrieval pipeline
Hybrid search, reranking and a citation check written in code, with a corpus you can trust and tests that never call the model. The pipeline degrades step by step instead of failing.
AI-assisted delivery for your team
Coding agents that plan before they build, with independent review and a merge gate a model can't talk its way past. Set up in your repository and taught to your engineers.
From the workshop
Built it. Wrote it down.
The methods behind the work, written down so your team can use them too.
Agentic Workflows: Building a RAG Pipeline That Cites Its Sources
Build the Howard's Lantern pipeline by hand, one stage at a time, in TypeScript, and compare each stage with the production code.
On Amazon
The 25x Developer
Building real applications with AI agents, quickly, without giving up the testing that keeps them working.
On AmazonThe person behind the work
I’m Joel.
I've built software for twenty years: web, mobile and data platforms, a fleet-inspection app at Zonar that I moved from Angular to React in 60 days, and cargo systems for U.S. Customs and Border Protection. Since 2024 I've built production AI systems as a contractor with Pilot & Pixel, and for my own products. I studied applied mathematics at George Washington University and work from Frostburg, Maryland.
Tell me what you're building
A short note on the product, what the model does in it, and what goes wrong when it's wrong. I'll reply within two working days with whether I can help and what it would take.
Or reach me on LinkedIn.