Now

A snapshot of what I'm actually doing this month. I refresh this page when something material changes, not every week.

Shipping

• PromptForge is live at promptforge.uk. Free during the evaluation phase. The latest pass replaced the template-assembly engine with a hybrid-retrieval RAG pipeline: pgvector inside Supabase, BM25 fusion via RRF, optional CrossEncoder rerank, and a server-side citation validator that strips any sentence the model can't tie back to a retrieved chunk. The eval scoreboard (recall, precision, citation accuracy) and per-component ablations live in the repo. The older guided wizard still works at /session/new.
• Hijama Business: trilingual booking site for a family-run clinic in East London. Locale-prefixed routes for English, Bangla, and Arabic, with full right-to-left support for the Arabic locale. Mid-build, deploying soon.

Working

• Part-time at TwelveTech Systems on customer-facing web work. API design, code review, QA, and AI-integration spikes inside products that already ship.

Reading

• RAG papers, and the rebuttals. The real question is when an LLM in the loop actually helps versus when it's just latency dressed up.
• Anthropic and OpenAI release notes, mostly the bits on prompt caching and structured-output reliability.
• Self-improving agent skills: how to make a workflow file that gets sharper after each session instead of fossilising on day one.

Open to

• Graduate software engineering roles starting Sept 2027 (final year ends June 2027).
• Freelance work that involves real AI integration, not LLM cosplay.
• A chat with anyone building dev tools, agent frameworks, or AI products with serious engineering rigour behind them.