Skip to content

Work

What I've been working on.

A few recent projects I can discuss publicly. Most of my work has lived inside startup products or under client confidentiality, so this isn't a complete portfolio.

Argon AI · Life sciences

Turning RAG into a system you could trust

I re-architected Argon's core research pipeline from linear RAG into an agentic, tool-calling system. It gathered evidence across source documents and produced analysis backed by excerpt-level citations. I also built the evaluation harness that measured citation coverage, accuracy, and reliability—so the team could prove whether a change improved the product instead of relying on a convincing demo.

Argon AI · Life sciences

Running grounded analysis at table scale

I built Matrix, a system that ran retrieval and LLM analysis independently across every cell in an analytical table. Each result carried its own reasoning and supporting citations, making large comparisons possible without losing traceability. I later added sandboxed execution for AI-generated analysis code and interactive quantitative visualizations.

Confidential client · Due-diligence questionnaire platform

Giving agents documents they could reason over

I built a document-processing pipeline that converted uploaded DOCX questionnaires into clean HTML for agent reasoning. That gave the system a consistent representation of each document for ingestion, indexing, retrieval, and response generation—rather than forcing agents to reason over an opaque office-file format.

Argon AI · Life sciences

Making the product around the AI dramatically faster

I led Argon's migration from Next.js Pages Router to App Router and reworked how the application loaded, rendered, and transferred data. The migration reduced first-load JavaScript by 70% and cut build time by 85% to under three minutes. A separate assistant refactor reduced render latency from 1.2 seconds to 400 milliseconds and cut response payloads in half.

The kind of work this usually is.

Engagements tend to fall into one of these four shapes.

AI product engineering

LLM features that hold up once strangers are using them. Retrieval, agents, and pipelines you can measure, debug, and keep running.

Full-stack product work

Features taken all the way: React on the front, typed APIs behind it, a data model that isn't fighting you. Including the unglamorous last 20%.

Architecture & modernization

For systems that outgrew their original shape. Making them faster and less fragile without betting the company on a rewrite.

Embedded engineering

A senior engineer in your standups who owns real work and pushes back in design reviews. Without a headcount req or three months of interviews.

Want to talk through yours?

Send me the shape of the problem and I'll tell you how I'd approach it, roughly what it takes, and whether it's a fit. Reach me at keaton@bytecraftco.com.