(systems in production)
what we've built
sales history database + natural language ai query layer
Queryable sales history where no analytical database existed: 900,000+ transaction records across 78 weekly snapshots, normalized for multipacks, packaging changes and retailer quirks. End-to-end ingestion, storage, domain logic, AI access. DuckDB + Parquet, Python pipeline, MCP server for plain-English queries.
amazon ads reporting pipeline
API report retrieval, Parquet snapshots as the source of truth, a rebuildable DuckDB store, console-export ingestion that normalizes exported columns. Read-only MCP query interface.
discovery intake agent (dia)
Discovery questionnaire in -> scored opportunity, client profile, opportunity analysis and proposal draft out. Deterministic parse, validate, score pipeline; a Gemini ADK agent calls the same functions as tools. Built for the Google for Startups AI Agents Challenge.
agent monitoring dashboard
Supabase-backed mission control for an AI agent fleet across VPS hosts: agent status, CPU/RAM, heartbeats, task board, pgvector memory search.
blackboard
Shared activity record for people and agents: handoffs, artifacts and events as plain JSON files, with a local viewer. Stays readable when the viewer is down.
baton
Agent-to-agent handoff over a private Tailscale mesh. One CLI any agent harness can call; every pass returns a trackable message id, and failed sends fail loudly.
bespoke
Operator-controlled voice layer: read text aloud, ask and hear the answer, live two-way audio with interruption handling.
content curation pipeline
AI agent that researches trending topics, curates archival content and drafts posts - connected to Obsidian + NotebookLM.
automated invoice processing
Document ingestion portal for SMBs and accounting firms. Upload invoices, get structured data - no manual entry.