Senior AI engineer · 10+ years in software
Most AI features die somewhere between the demo that got applause and the system that survives real users. I work on the second half — evals, retrieval quality, agent reliability, cost and latency — on products that are already live.
Engagements
Three ways to work together. Most people start with the audit, because it's the cheapest way to find out whether the rest is worth doing.
I go through your AI system the way an on-call engineer would, and hand back a ranked list of what will break and what it costs to fix.
I do the work. Fixing what's broken, or building the next version of the AI surface with the reliability layer in from the start.
Embedded with your team as the person who owns the AI surface area, so your engineers can own everything else.
Scope
Task-level success criteria, golden datasets from your real traffic, and CI that catches quality drift before your users do.
Chunking, reranking, hybrid search, and grounding that holds up past the few thousand documents where most RAG systems start hallucinating citations.
Bounded autonomy, retries, tool-call guardrails, and state persistence — so a multi-step agent fails safely instead of confidently.
Prompt caching, model routing, batching, and streaming. Usually the difference between unit economics that work and ones that don't.
Per-request visibility into what the model saw, what it called, what it cost, and where it went wrong. You can't fix what you can't see.
Full-stack when you need it: backend API, LLM layer, frontend, deployment. One person, one codebase, no coordination tax.
How it goes
Thirty minutes. You describe what's live and what's going wrong. I tell you whether I can help, what it would take, and what it would cost. If you don't need me, I'll say so.
One week, fixed price. I get read access, run evals against your real traffic, and come back with a ranked list of failure modes and fixes. Useful even if you stop there.
I work the fix list in scoped sprints, or embed part-time and own the AI roadmap. Deployed code, tests, docs, and a handoff call. Your repo, no lock-in.
Selected work
Podium
An AI agent that runs complete sales conversations for auto dealerships end to end, without a human in the loop, plus the customer data platform that gives it real context on every lead. Deployed on a platform serving more than 100,000 local businesses.
Caddi
A browser extension records how people actually do their work, then a video pipeline extracts key frames and combines them with transcription and LLMs to produce structured, executable workflows. Users query and modify their automations through a natural language chatbot.
Revado AI
Billing specialists query medical codes in plain English instead of writing SQL, cutting manual lookup effort by 40%. Also built the denied-claims dashboard, an automated denial-and-appeal workflow using speech recognition, and a PyTorch model that predicts claim rejections before they happen.
About
I'm Devin Jackson. I've been writing software for over a decade and working on AI systems full-time for the last two years. Right now I'm a Senior AI Engineer at Podium. Before that: Caddi, Revado AI, and Nickels, where I led a generative AI customer tool that drove a 20% engagement lift. Before the AI work, three and a half years as a senior engineer at LeafLink building a B2B e-commerce platform.
I started taking on outside work because I kept meeting founders in the same spot: an AI feature shipped, users on it, and no one on the team who had done this at scale before. The demo is the easy part now — writing the first version of a RAG pipeline takes an afternoon. Knowing why retrieval quality collapses at a hundred thousand documents, or why your agent's token spend went up tenfold, is the part that takes having done it.
I work remotely from Maryland, mostly with US startups between seed and Series A.
Next step
Thirty minutes, no pitch. I'll tell you straight whether this is a problem I can fix, roughly what it would take, and what it would cost.