Projects
AI, product, UX and technology projects exploring practical applications of intelligent systems.
Building useful things with technology and AI
Section titled “Building useful things with technology and AI”A collection of projects where I explore problems, design solutions and build working prototypes across AI, agentic systems, product, UX and software engineering.
I approach projects from both sides of the problem:
What is worth solving?
and
How can I build it?
Featured projects - AI & Agentic Systems
Section titled “Featured projects - AI & Agentic Systems”AI AGENTS · PRODUCT · AUTOMATION
AI Agent Job Scout
An agentic workflow that discovers opportunities, analyses job requirements, matches them against a candidate profile and helps prepare targeted applications.
Explore project →
RAG for Sherlock Holmes
Sherlock Holmes RAG Agent
A retrieval-augmented generation system over the Sherlock Holmes canon — ChromaDB, Ollama embeddings, semantic chunking, and a 30-question evaluation harness measuring hallucination rate, retrieval accuracy and multi-hop reasoning.
Explore project →
AI · CHILD SAFETY · GOVERNANCE
Habitat Hero Academy
An educational ecosystem-building game for children with an AI riddle generation layer — designed around layered child-safety guardrails, cost governance, offline-first usability, and an operator console with kill-switch and audit logging.
Explore project →
LLM · FINANCE · AUTOMATION
Financial Market Brief Generator
An AI system that generates audited daily market briefs from SEC filings and news. A deterministic data pipeline produces evidence bundles; a Researcher agent drafts claims with citations, a Critic agent reviews them, and a revision loop ensures every claim is grounded — or the brief gets flagged, never silently shipped.
Explore project →
Machine Learning & Data
Section titled “Machine Learning & Data”Experiments
Section titled “Experiments”Technical explorations where I test ideas, measure results and learn from failure.
Built a RAG agent over the Sherlock Holmes canon (13 documents, ~148K words) using ChromaDB and Ollama nomic-embed-text. Evaluated with a 30-question test set across four categories — achieved 0% hallucination rate and 57% correctness, with retrieval misses identified as the dominant failure mode. Lessons learned →
Built an agent harness using the Claude Agent SDK with an MCP server that auto-discovers tool modules — resume parser, semantic scorer, job search — and exposes them as callable tools in an interactive agent loop. See architecture →
Implemented a scoring engine combining 0.6 × cosine similarity + 0.4 × skill overlap using local embeddings (Ollama nomic-embed-text, 768-dim) against a persistent ChromaDB vector store, with configurable thresholds and seen-URL deduplication via SQLite. Tech stack →
Instrumented every LLM call with Langfuse for token usage, cost tracking and latency monitoring across resume tailoring, cover letter generation and matching — enabling per-stage cost analysis and performance debugging. Details →
How I build
Section titled “How I build”My projects generally follow the same principle:
Find a meaningful problem or opportunity.
Explore users, business needs and constraints.
Shape the product, workflow and experience.
Turn the concept into a working system.
Measure whether the solution actually works.
Reflect, iterate and improve.