Selected Work · 2024—2026

Things I've built.

Four projects across two chapters — a shipped product that proves I deliver, and the 2026 flagship slate I'm building now. Each is framed the way it should be: the problem first, the decisions second, the stack last.

◦Chapter I · ShippedReal products, in the world
01 · Full-stack · AI · Healthcare
Live

Mediot.

AI personal health companion
The problem

People get medical reports they can't read and don't know how to act on. Care advice is scattered, jargon-heavy, and rarely in their language.

What I built

A companion that scans reports with OCR, explains them in plain language, runs a symptom analyser, and answers follow-ups through a multilingual AI chatbot — so people can make informed decisions about their own care.

OCR
Report scanning
Multi-lang
i18next
AI chat
Gemini API
React.js Node · Express MongoDB Cloud Vision · Textract Tailwind
Visit mediot.vercel.app ↗
mediot.vercel.app
◦ ANALYZING REPORT
AI SUMMARY
◦Chapter II · The 2026 slateIn active development
02 · Flagship · In development

Atlas.

AI workflow & document intelligence

A platform where a team uploads their messy files and AI turns them into dashboards, answers and one automated report — so nobody copy-pastes spreadsheets again. It's what I do at BEC Chemicals, generalised into a product. Shipping as a thin slice first: upload → AI extraction & summary → auto dashboard → ask questions of your data.

Core
Upload CSV/PDF, AI extraction, RAG Q&A, auto dashboards.
Then
Roles & audit log, scheduled reports, automation builder.
Stack
Next.js, Node, Postgres, pgvector, Gemini/Claude, OCR.
03 · Agent · AI · Automation
In dev

Flux.

Autonomous ops agent
The problem

Reconciling documents by hand — matching purchase orders to deliveries, flagging mismatches — is slow, repetitive and easy to get wrong.

What I'm building

An agent that plans, calls tools, checks its own work and reports — with real guardrails (step caps, validation, human-in-loop) and an eval set that measures how often it's right. You watch it think in a live trace.

Plan→Act
Agent loop
Tool use
Schema-safe
Evals
Measured
Node · TS Claude · Gemini tools React · SSE
◦ AGENT TRACE
plan · 4 steps
call parse_document(po_112)
obs 8 line items
call compare_records()
flag qty mismatch · item 3
done report ready
04 · RAG · AI · Retrieval
In dev

Ledger.

Chat with your documents
The problem

Most document chatbots confidently make things up — no sources, no honesty about what they don't know. You can't trust the answer.

What I'm building

Upload PDFs, get answers that cite the exact page — and an honest \"I can't find that\" when the docs don't cover it. A guest sandbox lets anyone try it instantly, no signup. Its retrieval core becomes a module inside Atlas.

Citations
To the page
Honest
No hallucinations
Sandbox
Try instantly
React · Node pgvector Gemini embeddings
Q · what's the renewal date?
contract.pdf · p.4p.7

Want the full story on any of these?

Get in touch →