Production work rebuilt as demos with synthetic data, and public projects you can run yourself. Synthetic fixtures are labeled and never presented as real-world usage.
Built at Revotas, a B2B marketing and customer-experience SaaS. The code and client data belong to the company, so each case study recreates the product with synthetic data.
A back-office console written from scratch, with cited AI answers over internal documents
A back-office console I wrote from scratch as its only author. Staff run accounts, contracts, invoices, support and billing from one place.
One script that puts search and autocomplete on any storefront, live on 271 brands
A search overlay and autocomplete that attaches to a storefront's own search box with one script, configured per brand without code.
A CLI that took search from one brand to 271 without touching a storefront
A TypeScript CLI that installs Search V2 through the loader each site already runs, and dry-runs by default.
Recommendations rendered in the store's own card design and carousel
One config-driven script that copies a real product card from the page, fills it with recommendation data and starts the store's own slider on it.
A no-code builder that compiles a marketer's settings into a storefront script
Deal of the Day V2, built from scratch: a builder that turns settings into a self-contained storefront script, with a live preview.
Paste a product URL and the email block fills itself in
A block for the panel's drag-and-drop email builder that fills in a product's title, price, image and description from its URL.
A guard that stops saves landing on the wrong brand, and three years of panel work
Work inside the company's main marketing panel, a React codebase with years of history and many contributors.
Push notifications that open the right screen from any app state
The React Native side of push for the platform's mobile clients, from token registration to deep-link routing.
A support console where answers cite sources and a human approves every action
A TypeScript support operations console with cited streaming responses and a human approval gate. Seeded synthetic data.
One refund policy, two agent runtimes, same outcome: 50 lines against 97
One refund policy as a native tool-calling loop and as a LangGraph state graph, measured against a scripted deterministic model. Follow-up: the same refund decision from a chat model and a decision model.
Evals for citations, abstention, latency and cost that fail the build on a drop
Offline evaluation that produces the same result locally and in CI. No API key, no LLM judge.
Hybrid retrieval with rank fusion, citations and no-context abstention
A runnable retrieval reference. Fixture scores validate the contract, not real-world quality.
An AI-first platform for clinics. Private product, so there is no public demo.
Open Relay Desk, generate a cited reply, inspect proposed actions, approve a safe action, then review the synthetic reliability dashboard.
Open live demo ↗Review the golden-set contract, failure taxonomy, non-zero CI gate and deterministic HTML report in LLM Reliability Lab.
Open repository ↗Follow ingestion through lexical and vector ranking, RRF, citations, abstention and offline Hit@3/MRR fixtures.
Open repository ↗Open to full-stack and applied AI roles, remote or hybrid.
Tell me about the role, the team and the problem.