DEV Community

Cover image for Our AI factory built a safety product — and deliberately left the LLM out of it
Alex
Alex

Posted on

Our AI factory built a safety product — and deliberately left the LLM out of it

#ai #webdev #showdev #startup

Sentinel — an AI factory built a safety product and left the LLM out

Most factory demos end the same way: look, the agents wrote a CRUD app.

This one ends differently.

Sentinel is a live, embeddable weather / wildfire / flood advisory — invented by our AI factory after market research, written by AI directors, shipped by the pipeline, and deployed to Vercel. The twist is in the product brief:

No black-box AI at runtime. Deterministic thresholds. Signed evidence receipts.

The factory used AI to decide and build. It refused to use AI to judge safety.

Live: prod-bdb1634806de.vercel.app · Catalog: aicom-products

The brief came from the factory, not from a slide deck

Factory loop — analyst → marketer → ship

Inside AI-Factory, products move through a real pipeline: analyst → PM → marketing → architect → developer ↔ QA → devops. Sentinel is product id prod-bdb1634806de.

On 2026-08-15 the analyst agent filed market_research.json. The idea was already sharp:

Sentinel is an autonomous, **LLM-free* safety companion that any website can embed. It tells the visitor… what is happening right now with weather, wildfire and flooding — and it proves every statement with a signed evidence receipt instead of a generated opinion. Sentinel has no model of its own…*

The competitive gap it named:

Most competitors rely on AI or black-box models… Opportunity to own the *'verified AI-free safety' niche*.

Then marketing wrote the public voice — still from the factory, still on the record:

AI assistants are powerful, but they can hallucinate. When safety is on the line, a confident but wrong alert isn’t just embarrassing — **it’s a liability.

That sentence is the product.

Why the factory banned the LLM from the critical path

Why no LLM — liability quote from market research

Pain points from the same research file, almost verbatim:

  • Visitors need location-specific safety info; generic widgets ignore context.
  • AI chatbots may hallucinate alerts, exposing the business to liability.
  • Current tools make it impossible to prove what was said.

So the competitive advantage list starts with:

  1. Deterministic, published thresholds — no black-box AI reasoning.
  2. Cryptographically signed evidence receipts with digests.
  3. Micropayment metering on the AI-market protocol.
  4. Privacy-first: minimize location; don’t ship visitor data into an LLM.

This is the part people get wrong about “AI companies.” Using agents to design a product does not require putting a model behind every user-facing judgment. For safety, trust is an architecture choice.

What shipped

Sentinel live widget on Vercel

Sentinel Verified Safety Companion — FastAPI + React/Vite on Vercel fullstack.

Layer Choice
Runtime advisory Deterministic RuleEngine (no LLM)
Evidence ATLAS mesh invokes over AIMarket Hub
Caps atlas.situation.brief@v1 · atlas.fire.weather@v1 · atlas.nearest.read@v1
Levels CALM · WATCH · WARNING · EMERGENCY · UNKNOWN
Deploy prod-bdb1634806de.vercel.app · server: Vercel
Health GET /api/health → Sentinel Verified Safety Companion

One line embed (from the live page):

<script src="./api/embed.js" data-sentinel-key="YOUR_PUBLIC_KEY" async></script>
Enter fullscreen mode Exit fullscreen mode

Architecture — mesh invoke → rule engine → receipt

Flow:

  1. Visitor picks a location (rounded for privacy).
  2. Backend pays for ATLAS reads through the Hub.
  3. RuleEngine maps sensor payloads to published thresholds.
  4. Response carries a receipt digest when the mesh succeeds.
  5. If the mesh cannot be paid or observed — UNKNOWN, not a vibes-based “all clear.”

That last rule is not a bug. It is the thesis.

Honesty as a feature

Honest UNKNOWN when escrow cannot fund the mesh

I clicked Get safety report on the live deploy (Berlin coords). The UI returned:

  • Overall: insufficient balance
  • Weather / Wildfire / Flood: UNKNOWN
  • “How this was decided”: each layer not fired — no invented alert.

An LLM product would be tempted to smooth that over. Sentinel’s brief forbids it. If the factory cannot fund the sensor invoke, the product refuses to guess. Same spirit as our other “fail closed” work in the ecosystem — green UI is cosplay unless the numbers are measured.

What the factory actually did (timeline, condensed)

When What
2026-08-15 Product created · analyst → MARKET_RESEARCHED
Same day PM / marketing / methodologist / architect start · first CODE_COMMITTED
Mid–late Aug Long developer ↔ QA loop on mesh wiring (ATLAS invoke, not a fake constants advisory)
2026-08-28 Production alias on Vercel

Hundreds of factory tasks. Agents as marketer and directors. Humans as operators of the pipeline — reopen scripts, focus mode, escrow funding — not as the people who invented the niche in a workshop overnight.

Soft star / try it

If you care about AI that knows when not to speak, this is a concrete case study: agents found the market, agents wrote the copy, agents shipped the code — and the product they shipped does not use an LLM to call a flood.

⭐ if the “verified AI-free safety” niche is your kind of weird.

Every safety statement should come with a receipt.

Or it should stay silent.

Top comments (0)