Say hello Where ideas get potted
Small experiments in the open — built to show, not just tell. Sprout, my digital twin, floats in the bottom corner — tap the 🌱 to ask an AI version of me anything about my work. On the bench: the Writing Bed, which answers only from my posts, and Scout, an agent that ventures out to arXiv and Wikipedia. All three run on one tiny LangChain potting shed on Cloud Run.
Meet the residents
🌱 Sprout — the digital twin
A grounded chatbot that answers questions about me — roles, research, teaching — from a profile assembled out of the same data that renders this site. It speaks only from that profile: outside it, it says so and points you to my email. A small server holds the key, the persona and the rate limits; replies stream in token by token.
“What does Usama actually do?” — Sprout: “Usama teaches the full AI/ML lifecycle at Decoded, consults on GenAI systems, and publishes research — IEEE lead author. The CV has the full picture.”
How it's built: Growing Sprout
📚 The Writing Bed — answers from my posts
Retrieval-augmented generation over this blog. Your question is embedded as a vector, the nearest passages from my posts are retrieved, and the model answers from those passages — citing the posts it used. Ask about something I haven't written about, and it says the bed is freshly planted on that topic rather than improvising.
“Why did you rebuild your site?” — the Writing Bed: “The old FastAPI container said little about Usama and ran around the clock for four static pages; the rebuild is static-first with a brand built on ‘helping people and products grow’ — from Rebuilding Usama's corner of the internet.”
How it's built: The Writing Bed
🔭 Scout — the research agent
A small agent with two tools: arXiv and Wikipedia. It decides which to search — sometimes both, sometimes twice — reads what comes back, and repeats until it can answer with links to what it actually read. The loop is bounded by a step limit and a spend cap, so it always stops.
“What is retrieval-augmented generation?” — Scout searches, reads, then answers with sources: a definition grounded in the pages it fetched, plus links to the arXiv papers or Wikipedia entries it used.
How it's built: Scout
Answers come only from the posts on /writing/ — with receipts.
A live agent loop — so it's capped at 2 questions a day. Every answer cites what it actually read.
Let's grow something together.
Consulting, teaching, speaking, or a product idea that needs an AI brain — my inbox is open.
