I build the systems behind the feature: the data, the models, the checks, and the plumbing that lets a company rely on it.
Assistants grounded in your documents. Knowledge graphs of your expert domain. Agents with a person in the loop. Models you own. And the evaluation that proves any of it works — built, measured and handed over in a form your team can run.
I engineer the complete path from idea to production. That means the agent workflows that ground your AI in your own data, the evaluation loops that make its behaviour a number you can track, and the full stack around it: frontend, backend, database, deployment. You get a product your customers use, not a set of components.
Now A GraphRAG consulting platform and a speech-recognition stack for an LG subsidiary · a multi-agent marketing-campaign platform, active engagement · two engagements under NDA.
Your team asks in plain language; the assistant reads your own files, answers with the source shown, and can act on the answer — draft the reply, fill the form, open the ticket. When the source is not there, it says so.
When your knowledge is about how things relate — crops and treatments, parts and failures, rules and cases — a graph answers questions a document search cannot.
A team of AI agents does the research, scoring and drafting for one business workflow; a human approves before anything leaves the building.
Turn source material — recordings, PDFs, web data, a curriculum — into real PowerPoint, Word and Excel files that match your templates and open ready to edit.
When an API is too expensive, too slow, too inaccurate or not private enough, I fine-tune a model on your data and deploy it on your hardware — the training and the serving, both, for text and for speech.
“It doesn’t work well and we can’t tell why.” I turn that into numbers, then into an evaluation-and-optimisation loop that improves them release after release — the same loop that took my own models to state of the art.
Frontend, backend, database, deployment. No engineering team yet? I build the product around the AI as well, and hand it over in a form a team can take on later.
The company's agronomy know-how — guidelines, past reports, consultants' calls — answers farmers' questions with citations and drafts the consultants' reports for them.
Live in production · pilot expanding to all farms from May 2026
Consultants' phone calls become farming diaries and consulting reports. On real farming calls our fine-tuned engine makes 10.45% character errors; the leading commercial Korean speech API makes 16.39%.
10.45% error on real farming calls vs 16.39% for the leading commercial API
An English-test practice service needed instant, reliable scores for essays and spoken answers — at a scale no human grading or per-call vendor bill could support.
Nearly 1M users · supported a $5M+ raise
Teachers describe a lesson; the system returns an editable PowerPoint in their template, plus worksheets and quizzes drawn from a 120GB+ curriculum corpus.
Paying users; certified teachers use it in real classrooms
An exporter names a product; a crew of agents finds the overseas buyers, scores them on real trade records, and writes a report the sales team can act on.
Delivered March–June 2025, live at federationlabs.ai
An AI answer is split into individual claims; each claim is checked against the reference documents, and the answer is rewritten around exactly the claims that failed.
5★ review · $18K engagement
A marketing-campaign platform where AI drafts the strategy and the outreach, but an expert and then the client approve before anything moves — and a crashed run picks up where it stopped.
20 specialist subagents with two sequential approval gates
I plan, run and publish model research — the essay-scoring model that reached state of the art in its paper went straight into a million-user product. The habit of measuring before shipping comes from the same place.
State-of-the-art result, arXiv 2309.02740 — planned, run and written by meTwo recent engagements are under non-disclosure agreements. They are described here without the client, the industry or any figures — only what was built, so you can judge whether the capability fits your problem.
An AI companion that remembers the user across conversations and stays inside strict behavioural limits — including when parts of the system are unavailable, where it degrades gracefully instead of failing.
The AI “brain” and the client's platform are separated by a written interface contract with a stub mode, so their engineers integrate against a stable surface while the AI evolves behind it.
You describe what you want to happen in the business. I ask until I can say it back in one sentence. No slides yet.
Every item is a row: what it does for you, whether it already exists, hours, and price. Three stacking packs so you choose the size, not the vendor.
Each milestone has a written test you run yourself. Payment follows the test, not the calendar.
Decisions, deviations and progress live in a project document repository you can read at any time. Technical detail stays folded under plain-language summaries.
Each delivery comes with a report in your words: current state, what changed against the plan, what to keep in mind — and a checklist you tick off.
Code, infrastructure, evaluation sets and method notes. Vendor work is called “integrate”, never hidden as “built”.
Grounded answers, citations, your documents, a test set.
Evaluation loop, hard cases, the second workflow.
Login, roles, deployment on your infrastructure.
| Item | Status | Pack | Hours |
|---|---|---|---|
| Document ingestion, your formats | Proven | 1 | 12 |
| Cited answers with refusal | Proven | 1 | 18 |
| Your CRM as a tool the agent can use | Build | 2 | 16 |
Things I have built before cost you their integration hours, and the quote says which rows those are. You pay for the outcome, not for me relearning.
In the systems I build, a sentence without evidence is removed. In my proposals, a number without a source is not written.
A second pass, in a separate process, with a different model. It can flag and demote. It can never promote — so it cannot hide a failure.
Features that measure zero gain are switched off and documented, even after the work is done. The evaluation set is part of the handover.
Every scope lists what is not included. If a request needs something I have not done — AI video, no-code automation tools — I say so before you pay.
Thirty minutes. You leave with the problem stated in one sentence and a view on whether AI is the right tool for it — and if not, what is.