Building in the open.

A self-running weekly brief that ranks GitHub's trending repos by momentum, not lifetime stars, drops anything below a proven-star floor, then has an LLM explain each one in product terms: what it does, who it's for, and the week's theme. Built to answer a product question, not an engineering one.

Give it a topic and six small agents plan the searches, write a cited report, grade their own work, and re-research the weak spots before emailing the result. A compact, readable example of an agentic workflow with a self-correction loop and human-in-the-loop checkpoints.

A facts-only mutual-fund FAQ assistant, built in phases: it scrapes and normalises scheme pages, indexes them into Chroma with BGE embeddings, answers only from what it retrieves, refuses anything out of scope, and holds itself to a golden-set evaluation suite with daily re-indexing.

A manager agent delegates to a team of researchers, keeps persistent memory across runs, scans a sector for companies gaining traction, and picks one to back with a written rationale for the call and the rejections.

A pipeline that turns app-store reviews into a weekly product-signal report, from raw fetch through a schema-validated feed a dashboard can read.

An MCP-style server that lets an AI append to Google Docs and draft Gmail through structured, OAuth-gated tools, with a human approval step before anything runs.

A CrewAI crew where one agent gathers current facts on a company and another writes a structured financial report from them, passing context down the line.

A small CrewAI project that stages a debate: one agent argues for a motion, one against, and an impartial judge decides which side was more convincing. Agents and tasks live in plain YAML.

AI-assisted restaurant discovery: deterministic SQLite retrieval builds a grounded candidate list, then a Groq LLM re-ranks and writes short explanations for the shortlist. FastAPI backend, Vite front end, CI and deploy wired up.