Inspiration
Every big chip story this year has the same twist: the trouble didn't start at NVIDIA, Micron or Intel. It started a few steps down, at a company most people have never heard of. A substrate maker running short. A union dispute at a memory plant. A film only one company in the world makes.
Procurement teams and investors watch the giants, but the shocks come from further down the chain. The tools that map those hidden suppliers are expensive and updated by hand, so they're always behind the news. We wanted something you could just type a product into and get the answer, live, with proof.
What it does
Type a product or ticker, like Intel Xeon or Micron HBM3E. In about a minute, Deepchain:
- Maps the supply chain, from raw materials up to the finished product.
- Shows proof for every link: a quote from a real source, checked against the page.
- Watches the weak spots with Telem monitor agents and Tavily news.
- Turns a supplier red when there's real trouble, and shows everything that depends on it.
How we built it
- Telem search routes each question (filings, teardowns, supplier leads) to the best search engines automatically.
- Telem fetch turns web pages into clean text.
- Tavily brings in this month's news, a weekly headline pulse for every supplier, and a one-click "check for problems now".
- Claude reads the pages, works out who supplies whom, and suggests likely chokepoints for a second round of targeted searches.
- Telem Monitor Agents watch each critical supplier daily, with their own plan and ru plain English.
- Next.js for the app, hosted on Railway.
A supplier only turns red when a news event passes every check:
$$ \text{problem} = \text{about the supplier} \land \text{real impact} \land \text{quote fouustworthy source} \land (\text{age} \le 30\ \text{days}) $$
Once supplier $s$ is flagged, everything downstream of it is marked at risk:
$$ \text{AtRisk}(s) = {\, v \mid s \rightarrow \dots \rightarrow v \,} $$
Challenges we ran into
- Keeping the AI honest. It's easy to generate a supplier map that looks right but is made up. So the AI has to quote the exact sentence behind every link, and our code checks it's really on the page. No quote, no link.
- Thin maps. General pages rarely say "X supplies Y", so our first maps had only one or two links. A second round of targeted searches, one per likely chokepoint, made them much richer.
- False alarms. Our first monitors flagged good news, like a supplier winning big orders, as a problem. We rewrote how monitors are set up so they read all the news but only flag real disruptions.
- Trustworthy sources. Early news checks mostly found aggregator sites, which the system rightly rejected. Better search wording and Tavily's deeper search brought in outlets like Bloomberg and BusinessKorea.
Accomplishments that we're proud of
- Real red flags, not staged ones. Our demo shows actual current events: Ibiden's substrate shortage hitting Intel's Xeon chips, and a labor dispute at Micron Taiwan putting HBM memory at risk.
- Every claim is clickable. You can go from any link on the map straight to the sentence that proves it.
- The system says no. It visibly rejects weak sources and good news, which is why itsing.
- A full working product in one day: live search, maps, monitors, news checks, sign-up, and a hosted app anyone can try.
What we learned
- Search quality matters more than model quality. The same AI gives far better answers when it's fed better sources.
- "Show your work" is the feature. People trust a red flag when they can read the sentenc
- Live data is messy, and the same search can return different results minutes apart. Checks, retries and saved results made the app reliable.
What's next for Deepchain
- Let companies upload their own bill of materials and watch every part automatically.
- Add shipping and trade data to find supplier links that never make the news.
- Send alerts to Slack and email the moment a supplier turns red.
- Go beyond chips, to EV batteries, pharma and aerospace, where the same hidden-supplier
Built With
- ai-agents
- claude
- exa
- github
- llm
- next.js
- node.js
- perplexity
- railway
- react
- tailwindcss
- tavily
- telem
- typescript
- web-search
- zod
