Policy Canary
Policy Canary watches the FDA for small consumer brands and tells them which of their own products a new rule actually affects, with the deadline and what to change. The companies that need it, a supplement maker or a cosmetics line, can't read every regulatory notice, so the work of figuring out "does this touch us?" was being done by hand, in spreadsheets, every time the FDA moved.
I designed and built it, with a multi-model LLM pipeline doing the matching and analysis at its core.
The gap I built into
Regulatory monitoring splits into two useless extremes. At one end, free Federal Register alerts and industry newsletters ("here's what happened in supplements this week") that leave you to figure out whether any of it touches your products. At the other, enterprise intelligence platforms at $25K–$200K a year, built for pharma regulatory teams, not a 12-person supplement brand.
Nobody watched the FDA for YOUR specific products at a price a small brand would actually pay. That's the gap: the difference between "what happened" and "what does this mean for my products" is the whole job, and it was still being done manually, in spreadsheets, every time the FDA moved.
Tell it what you sell
Onboarding starts with what you actually make. Pick a product type, and supplements auto-populate from a 214K-product reference database with structured ingredients; anything off-database, you photograph the label and a vision model extracts them. You confirm the ingredient list to watch, and can add the manufacturer's FDA establishment identifier so a facility's troubles surface too.
By the end the system knows the exact formula of every product you sell. That's the whole basis for matching: nothing can tell you a change affects your product until it knows what's in it.
Then it tells you what's yours
Twice a day, a pipeline ingests every new FDA action across seven sources: Federal Register, openFDA recalls, warning letters, guidance documents, import alerts, and more. Each item is analyzed, tagged, and scored against every subscriber's product profiles.
Every product carries a plain status (all clear, watching, or needs attention), and when something matches you get the analysis on YOUR product, not the industry: what's affected, why it matched, the deadline, and what to do. Nothing relevant is hidden; nothing irrelevant is padded.
A multi-model LLM pipeline
The intelligence is a pipeline of models, each chosen for what it's good at rather than one model doing everything. Gemini Flash and Pro do the bulk enrichment at ingest: reading each regulatory item once and extracting a plain-English summary, the action type, any deadline, the affected ingredients and product categories, action items, and cited regulations in a single structured pass.
A second Gemini Pro reasoning pass handles the part I'm proudest of: cross-reference inference. Using a 950K-code substance database, it reasons about risk that transfers across categories (an ingredient flagged in one sector quietly affecting products in another) and tags those inferred signals distinctly from the direct ones. Claude Sonnet does the writing wherever quality matters: composing each subscriber's briefing in a real editorial voice, and the urgent alerts. A lightweight model classifies products and renders the final product-relevance verdicts, and OpenAI embeddings power search. Three providers, each carrying the load it's best at.
Automations run the whole thing
It's built to run without me. A cron orchestrator fires the daily ingest (seven fetchers in parallel, then enrichment) and the weekly send, fanning a personalized briefing out to every subscriber. Onboarding, matching, alerting, billing, and analytics are all wired end to end.
The marketing runs itself too. A self-hosted AI content agent queries the same enriched database, drafts SEO posts and weekly roundups on a schedule, routes them to me for a one-word approval, and publishes to the blog. The product's own data pipeline is the engine for the content that sells it.
AI as the team
Policy Canary was the next rung: the same way of working aimed at something much harder. Not a directory but a production SaaS where the analysis has to be right, running on a multi-model pipeline that reasons across a 950K-code substance database and hands real companies decisions they act on.
This is where "AI as the team" got literal. The pipeline does the regulatory analysis a team of specialists would; the automations run the ingest, the sending, even the marketing; and I designed and built the whole product around them. It's the clearest case of using AI to ship a real business's worth of software alone.