I build agents, retrieval systems and the evaluation harnesses that keep them honest — from a triage agent that knows when not to reply, to browser personas that find, verify and fix bugs on their own. Rigorous where it counts, playful where it pays.
A triage agent for a tax authority's inbox. Emails are PII-scrubbed, classified by a fine-tuned MiniLM into 10 intents, calibrated, and either drafted with Gemini or escalated to an officer. The point is knowing when not to draft — in LangGraph, escalated emails have no path to the drafting node.
A self-healing loop for web apps. Five LLM-driven browser personas — flow, pricing, contract, security, data integrity — use a shopping app like real users and flag inconsistencies. A verifier confirms bugs, a sandboxed agent patches them with regression tests, and a dashboard replays every run.
Turns prose into a persistent, explorable 3D world. Each passage patches the world instead of making a fresh scene, so the desk in chapter three is the same desk from chapter one. I built the narrative engine that extracts places, objects and relations from the text.
An agent that analyses an investment portfolio end-to-end: risk metrics, a neural-net risk classifier and an LSTM volatility forecast, explained through a 4-base RAG knowledge store — with a hallucination guard that checks every number.
A small black cat that lives on your desktop. She mostly just watches, but when you're reading a PDF or replying to email, she offers a quick summary or draft reply. A seven-sprite state machine, persistent memory and optional voice, built in a 3-hour hackathon.
Turns books into short, opinionated video essays — one episode per law or chapter. A staged pipeline from script to video, with a six-beat narrative skeleton, a persona-driven host and a tier list that updates across the season.
A BuzzFeed-style quiz platform that generates quizzes on the fly from Google Trends topics, with Gemini-written questions, Pollinations.ai cover art and "X% of people got this result" stats.
Model comparisons, retrieval pipelines and datathon builds — where I test ideas before they become products.
Hybrid retrieval (FAISS + BM25, fused with RRF) over FOMC minutes and statements from 2020–2025, feeding a 4-bit Mistral 7B — evaluated with ROUGE and a Claude LLM-as-judge.
FinBERT vs RoBERTa-large on hawkish/dovish stance classification — and why domain pre-training lost to the bigger general model (F1 0.71).
Segments 8.5k companies with K-Means, flags outliers with Isolation Forest, benchmarks IT spend against peers, and writes executive briefings with a fine-tuned FLAN-T5.
Pairs XGBoost/ARIMA forecasts with DCC-GARCH and LSTM covariance models inside a constrained optimiser, rebalanced weekly across 8 coins.
The packages behind the projects above — taken from each repo's requirements.txt and package.json, with where I've used them.
I build AI systems that are useful, measurable, and honest about what they don't know.
Most of my recent work is agents and the scaffolding around them: routing that escalates instead of guessing, retrieval that cites its sources, browser personas that test software like real users, and evaluations — calibration, risk-coverage curves, human review — that tell you whether any of it actually works. I care a lot about the unglamorous parts: safety gates, reproducibility, and benchmarks you can trust.
But I also build for the joy of it. A desktop cat that summarises your PDFs. A quiz generator that runs on trending topics. A pipeline that turns books into video essays. Dashboards and sites I can ship in a weekend. The playful projects keep the rigorous ones honest.