Projects & skills
Things I build outside my day job, from course prototypes to products people use.
Technical lead, for Kerubin Capital with Investing in Women and the Manila Angel Investors Network. There wasn't one place to look up women-led and gender-lens ventures in the Philippines, or the people funding them. I built a directory and network map to fix that. It runs on Next.js and Vercel with Airtable behind it. Four types of organisation can submit their own entries, and the map shows who is connected to whom.
Co-founder and CTO. Most people book a wildlife trip to see one animal, and whether they do depends a lot on the guide. Lyferr is a marketplace that starts from the species. It matches travellers with local guides who know that animal in that place, and puts bookings next to guide expertise and real-time scientific data. I took it from concept to a live platform in five weeks and designed the two algorithms behind it.
Built with Henry Kurze (Hertie School) and Laura Luca (SNSPA). EU rules like the Digital Services Act ask platforms to be open about what they host. Most auditing tools still work on text, while more and more of what people watch is short-form video. We built an auditing tool that takes a video and checks it against independent fact-checks and EU platform law, comparing it with the moderation logs platforms publish about themselves.
Set up as a chain of specialised models: Whisper transcribes the audio, Hive classifies what's on screen, ChatGPT and Claude cross-reference the result against the legal frameworks and draft the policy notes.
For years I watched the President's State of the Nation Address with a notes app open, logging which topics came up and timing how many minutes each one got. This year I turned my manual tracking into a weighted scoring model. Claude handled the volume: six months of pre-SONA news, investment pledges, priority bills and coalition stances, scored and written up.
Care homes run on rotating staff. Between agency workers, shift swaps and new hires, much of what carers know about a resident is lost at handover. I built a prototype RAG assistant that gives care staff short refreshers when they need them — answering only from eight public clinical guidelines, never making anything up.
I split the guidelines into 594 passages, each tagged with source, date, language and topic. A small open model decides whether a question needs a search, what to search for, and whether to narrow the search by date or source. The search filters before it ranks, so a request for recent guidance never returns a 2009 document. When no passage clears a relevance threshold, the assistant says it doesn't know — the model is never asked.
What didn't work at first.The first version made up an answer. A stricter prompt didn't stop it, so I enforced the rule in code: hard relevance thresholds and a fallback re-search without the filter. I also ran evals across the full guideline set to find where retrieval broke down and to compare the 1.5B vs 3B model on answer quality.
Eight sections laying out how I'd approach the role: why HappyRobot and why now, two past partnerships and what I changed after each, how I'd work alongside an FDE, a full account prep for a first telecom customer from Madrid, a measurement framework, and a 30-day plan.
Skills
- —Retrieval-augmented generation (RAG)
- —Embeddings & vector search
- —Evaluation (evals)
- —Agent routing
- —Multi-model pipelines
- —Gemini API
- —Whisper · Hive · Claude · ChatGPT
- —Python (pandas, PyTorch, Hugging Face)
- —FAISS
- —SQL & BigQuery
- —TypeScript
- —C# / ASP.NET
- —Next.js & Vercel
- —Airtable
- —Google Cloud
- —Figma

