RAG pipelines, LLM-powered tools and memory layers for AI — custom-built around a specific problem instead of bolted on as a generic wrapper. I care about the version of software that actually ships.
An agentic cold-outreach engine that researches before it writes. It scrapes a lead's own site, retrieves the most relevant passages with RAG, and has Gemini draft a personalised email grounded in that context — scoring its own confidence and routing anything uncertain to human review instead of guessing.
A cross-platform memory layer for AI chat. It reads a conversation straight off the page, compresses it through a RAG pipeline into a clean summary, and injects that context into a brand-new chat on a different platform — so Claude can pick up exactly where ChatGPT left off.
An intraday trading simulator implementing a gap-up rejection strategy built from three years of personal market observation and manual trading. Sub-millisecond latency per candle, constant space complexity O(1). Capital-based 2% stop-loss / 7% target risk model, single position, max two trades a day, automatic 15:00 square-off.
Most AI products are a thin wrapper around an API call. I build the part underneath: retrieval that grounds a model in real data, memory that follows you between tools, and agents that know when they are unsure.
I work where retrieval, language models and the web meet. Away from the keyboard I read and write — mostly poetry, occasionally a novel I'm slowly working through.
Open to new roles and interesting problems. Reach out to me directly: