I ship LLM systems to real clients, and build companies around them.

Vantaphai

Your digital knowledge lawyer.

Law firms track regulatory change by hand across dozens of fragmented sources, and things get missed. Vantaphai continuously scans legislation, court decisions and regulator publications, then turns them into risk-weighted alerts and obligation timelines tailored to each client and practice area.

I built the full platform through paired programming to the point of deployment and pilot, built the landing page, and work directly with large bank, VC and law firm clients to shape what gets built next.

• Retrieval-augmented generation with a Cohere cross-encoder reranker for legal document analysis, cutting input tokens on large statutes by roughly 55 to 60% against the capped baseline
• LLM evaluation harnesses for relevance, judge consistency and tagging: judge consistency to 20/20 on synthetic cases, tagging output tokens down about 40% at the same recall
• Planner-routed agentic tool-calling, and a novel knowledge-graph representation of regulatory and document corpora
• CI/CD via GitHub Actions and code review throughout, with mentorship on secure AI system design from a former DeepMind specialist

• Python
• TypeScript
• Next.js
• RAG
• Cross-encoder reranking
• LLM evaluation
• Knowledge graphs
• GitHub Actions

vantaphai.com ↗

Supercharged

Find internships through the people already there.

Cold outreach is how students actually land spring weeks and graduate roles in finance, trading and consulting, and almost nobody does it well. Tell Supercharged the role, the firms and the season; it finds the alumni and analysts already there, scores who will actually reply, and drafts the email in your voice. Then it tracks the whole pipeline.

• Multi-step agentic search: a planner routes between web-search strategies, external APIs, user clarification and vector search over a cached network (Gemini embeddings, Convex vector index)
• Multi-agent LLM-as-judge using two model families to verify and score matches, avoiding agreement bias
• 291 on the waitlist; October pilots lined up with the Cambridge Marshall Society and Oxford Alphafund Society
• Student ambassadors across seven universities; shipped via GitHub Actions with automated testing

• Next.js 16
• React 19
• TypeScript
• Convex
• Clerk
• Gemini
• PostHog
• Playwright

superchargedai.app ↗ Marketing site source ↗ Oxford launch MVP ↗

The product codebase is proprietary. The marketing site and the original MVP are open.

Supercharged: Grow Together (v1, codename Scrollwise)

The feed you're addicted to, pointed at things worth reading.

A React Native app that keeps the mechanics of a short-form feed but fills it with AI industry summaries and long-form ideas: articles, podcasts and books, surfaced by search rather than an engagement optimiser. Released on the iOS App Store in September 2025 as the first product under the Supercharged name.

• Hybrid retrieval: Gemini embeddings in pgvector and Postgres full-text search, fused with Reciprocal Rank Fusion
• Progressive search returns lexical hits instantly, then re-ranks once the embedding round-trip resolves
• 13 Deno edge functions, roughly 160 components, cron-batched push notifications with receipt pruning

• React Native
• Expo
• TypeScript
• Supabase
• Postgres
• pgvector
• Deno
• Python
• EAS

Source & write-up ↗

Get There One Day

A 100% Q&A podcast for ambitious young people, and the community that grew around it.

Answers the questions degree apprentices and school leavers actually ask. It passed 16,000 followers in four months, partners with brands on sponsored content, and got me invited to speak at the Independent Higher Education conference in November on young people's changing attitudes to education.

I also built the site: a free Degree Apprenticeship Playbook, and Charge, an AI assistant that answers apprenticeship questions and reviews CVs and cover letters, with a paid tier.

• Charge streams replies live from a Convex backend, with per-user daily caps and monthly cost caps
• Uploads are parsed in the browser so only text ever reaches the server, and it is purged after 30 days
• Stripe subscriptions, Clerk auth, free and £10/month tiers on different model providers

• React
• Vite
• Convex
• Clerk
• Stripe
• TikTok

TikTok ↗ Site & Charge source ↗

Athenalearn

One-to-one tutoring from students at top universities.

A tutoring startup I ran through my first year at Bath. I recruited and managed a team of more than ten tutors, interviewing each one, and personally delivered over 200 hours of one-to-one tutoring across seven clients. On the technical side I built the website and the interactive progress reports parents used to see how their child was doing.

• 10+ tutors onboarded and managed; 200+ hours delivered personally
• Interactive progress reports built on Python analytics
• Pantheon, a knowledge-management mobile app prototype in React Native

• React
• Vite
• Python
• React Native

Cortex

A personal AI brain that lives in a folder of markdown.

A local web app that puts a streaming chat interface on top of a plain folder tree. The agent, built on the Claude Agent SDK, reads and writes that tree directly: knowledge notes, its own persistent memory, uploaded files and skills. Because it follows Claude Code's conventions, the CLI and the web app share one brain and one chat history.

• Express and WebSocket server that spawns the agent per message with the brain as its working directory
• React front end with a file browser, drag-and-drop inbox and live-streamed replies
• The brain is a git repo, so every edit the agent makes is diffable and revertible

• TypeScript
• React
• Express
• WebSockets
• Claude Agent SDK

Source ↗