Bring me a process, a prototype or a pile of PDFs. You get a working system: LLM extraction pipelines, agents, RAG, multi-tenant SaaS. The first step is always the same: a vision document, an hour-by-hour estimate and a demo video, before anyone commits a budget.
Most of my work starts with something that is done by hand, by a chat prototype, or not at all.
PDF catalogs, scans, Excel with merged cells, portals without an API. An LLM pipeline reads them as they are, keeps values verbatim, flags what a human should check, and exports clean records.
A chat demo becomes a multi-tenant SaaS: user terminals, staff workstations, admin back-office, billing, audit. Delivered in increments, in production, with real users on it.
Telegram bots, Google Sheets backends, browser agents that walk a portal like a person. Small systems that remove hours of manual work every week for a small business.
Client names are omitted; the numbers are real. Two of the projects have a short walkthrough video.
A healthcare services provider had a single-tenant chat prototype for guided consultations: no clinician workstation, no billing records.
I designed and built a multi-tenant platform: a patient terminal (intake, safety questions, photos), a clinician workstation (queue, case page, approve/override) and an admin back-office (tenants, fees, invoices, audit).
A furniture marketplace onboarded a vendor a week, ~1,000 SKUs each, as PDF lookbooks, scans and Excel price lists with photos in merged cells. An operator spent 2–4 days per vendor retyping.
I built a Python pipeline that reads the files as they are: one LLM call per file for the layout, geometry for photo-to-product links, verbatim values, structural flags, a review screen, export as drafts.
Law firms request medical records through portals that have no API. Staff log in, search, click through forms and download files by hand, dozens of times a day.
I scoped and demoed an agent that runs the portal in a real browser: a scripted scenario does the clicking, the LLM maps each request to the scenario's options, every step is logged and visible in a watch console.
Shoppers ask in plain language: “a grey sofa under 800 for a small living room”. Classic filters miss it, pure vector search returns noise.
I built a hybrid router that decides per query between structured filters and semantic search, added image search, made chat answers deterministic, and wrote the tooling that keeps the vector index in sync with the catalog.
A real-estate consultant sold service packages by voice and kept everything in spreadsheets. Every client meant retyping the same questionnaire and building the same documents by hand.
I built a Telegram bot that walks the client through a 25-question intake, turns answers into a lead card and the consultant's own Excel templates, posts them to the team group, and a small web panel for managers. Runs on a VPS, webhook behind Caddy.
Before anyone commits a budget, you get three things. They are useful even if we never work together.
Chosen for boring reliability, not novelty. I pick the simplest thing that survives production.
Telegram is the fastest. I reply the same day, usually within a couple of hours.