Work · Case studies

Work that
shipped.

Warehouses, semantic layers and dashboards, owned solo — and the LLM systems on top of them, built so their answers can be checked rather than taken on trust.

2026 ● FEATURED

lease-validator — evidence or nothing

An agentic RAG engine that checks lease agreements against YAML rules and refuses to say anything it cannot quote the document for.

Python
Anthropic API
FastAPI
Pydantic

2026 ● FEATURED

metrics-mcp — a semantic layer an agent has to go through

Metrics defined once in YAML, SQL built from dbt, and every number an agent states checked against the result it came from before the sentence is allowed out.

Python
MCP
dbt
DuckDB

2025 — present ● FEATURED

UrbanStack — the analytics function, from zero

Sole data hire at a proptech SaaS: ingestion, a 124-model dbt warehouse, a 155-metric registry, four dashboards that replaced Looker Studio, and the models and alerts on top.

BigQuery
dbt
Python
Next.js

2026 — present

Skriptly — transcription that knows when it is wrong

A speech-to-text product built solo: serverless GPU workers, diarisation merged by word-level alignment, an LLM correction layer, and a measured error rate rather than a claimed one.

Python
Flask
Modal
faster-whisper

2025–2026

F1 Dashboard

A three-page Streamlit dashboard for Formula 1 data — lap times, qualifying breakdowns, and full season standings. Built with FastF1 and the Jolpica API.

Python
FastF1
Streamlit
Plotly