SiliconMatch

Overview

SiliconMatch helps a company that sells a chip, or an all-in-one box built around one, find the companies that need it. It searches a base of 1,721 verified prospect companies (as of 21 October 2026), scores each one for your offer with reasons that cite real sources, shows the people to approach and drafts a first message and follow-ups. You keep the good ones in lists and export them.

Version 2, the selling desk, adds: a dossier per company (filings, papers, news, timeline), a demand map with the sites prospects operate, signals with a cited “why now”, watchlists and an inbox, market trends and forecasts, a knowledge graph of who builds on whom, a practice room that rehearses your pitch against simulated buyers, a morning brief read aloud, and voice commands. Version 1 stays available at /v1/.

Demo sign-ins

PersonaPINRoleWhat you can do
VoltEdge Silicon (fictional)1111Seller of motor-control chipsSearch and score, dossiers, map, trends, graph, practice room, voice, the morning brief, masked contacts, sequences, own lists and watchlists
NeuroBox Systems (fictional)2222Seller of an edge-AI boxThe same, for its own offer, lists and brief

Each persona sees only its own lists; the server enforces this. The search form starts with the persona's own offer filled in.

User manual

1. Sign in

Open the live demo and enter a PIN from the table above.

2. Describe your offer

On Search, pick your chip or solution type (type in the filter box to narrow the list). If you sell a box, tick its parts: optimized hardware, local AI runtime and models, domain skills and agents, MCP and protocol connectors, test bench, observability.

3. Choose who to find

Tick groups or single segments. They come in three kinds: they build with chips (drones, robots, OEMs, ODMs and EMS, module makers), they run it in operations (manufacturing and test labs, energy and telecom, healthcare and labs, automotive and logistics) and they install it for others (integrators). Picking a chip type pre-ticks the segments it usually sells to. Add a country if you want.

4. Find prospects

Press Find prospects. A progress bar shows the work: choosing candidates from the base, then scoring the most promising ones for your offer and checking for competitors. About 60 companies take around half a minute.

5. Read the results

Results are sorted by fit. A green score means at least one reason is backed by direct evidence of need. Filter by role, sort by name or country, or show direct evidence only. Details opens the reasons with their sources, the opener, the integration story (for operators and integrators) and the roles to approach.

6. People and sequences

Under People you see the decision-makers we hold for that company, ranked by role. Draft sequence writes three e-mails (opener, follow-up with one proof point from your offer, polite close). Edit any step and save it. Nothing is sent from SiliconMatch.

7. Lists and export

Pick a list (or + New list and a name) in the results bar, then Save companies into it. On Lists, set each prospect's status (new, contacted, replied, meeting, not a fit), keep notes, open people and sequences, and use Export CSV. The file has one row per contact (or per company when it has none) with company, website, segment, country, fit, reasons, evidence links, pitch, contact and sequence.

8. Dossiers

Click a company name anywhere to open its dossier: identity and country (with the quote that states it), fit for each offer you have scored, a dated “why now” when a recent signal applies, news and signals, SEC filing passages, research papers, every piece of evidence grouped by source, people (masked in the public demo) and a timeline. PDF brief prints a one-page summary.

9. Map

Map shows prospects for an offer: a pin where a company's own sources name its city, otherwise shading for its country. Layers: competitors, “need stated on their own pages”, and Sites (power plants, hospitals and factories the prospects operate, each with its registry and licence). Click a pin or a country for the companies behind it.

10. Inbox and watchlists

Press Watch on a result, or on Inbox watch a whole segment or country. New signals for those companies (news, funding, expansions, contracts and tenders, hardware job posts) appear in the inbox, newest first, with links. Mark all read clears it.

11. Trends

Trends opens with the markets “heating up”: forecast growth for the next six months with its range, company moves in the last 30 days, and the three strongest moves behind each. Below are the eleven monthly series with their sources; a dashed line and shaded range show a forecast where one is shown.

12. Graph

Graph answers path questions (which integrators reach hospitals, who is exposed if NVIDIA changes its modules, who speaks factory protocols, Blue UAS makers, grid operators) or shows any company's links. Click a link to read the quote it comes from.

13. Practice room

Practice rehearses a pitch, compares two messages, or plays out a rival's price cut. Choose how many simulated buyers (and competitors) to use; results show the split of answers, the main objections, what would change each buyer's mind, and the facts each one relied on. Everything there is a hypothesis.

14. Morning brief

Brief plays a 60–90 second summary read by Ellen: what changed for your watched companies, five accounts to call and why now, and the market to watch. The facts it used are listed under the transcript.

15. Voice and typed commands

Click the microphone, say a command, and click again; or type it into the box beside it. Examples: “show agricultural drone makers in Brazil with direct evidence”, “open the dossier for Locus Robotics”, “who is affected if NVIDIA changes its edge modules?”, “watch Agility Robotics”, “play my brief”. If a name is unclear the app asks which one you meant.

Data and evidence

Sources. (1) Web search: Claude Haiku searches the web for each segment and extracts the companies the results name; an evidence link must be one the search actually returned. (2) Each company's own homepage: its title and description are stored as evidence when we verify the site. (3) The US FAA's accepted drone Declarations of Compliance (507 at the time of writing). (4) The US Department of Defense Blue UAS Cleared List (drones) and Framework List (components).

Verified websites. A company counts only if its website loads and shows the company's name. List sites, directories and market-report publishers are never accepted as a company's own website.

Countries. One English spelling per country. A country is shown when the evidence states where the company is based (the quote is kept and shown on hover), or, marked “(domain)”, when the website uses a country domain such as .de or .co.uk. Otherwise it is left blank rather than guessed.

Roles. Builder (makes end products), module maker (designs boards and components, so buys chips too), adopter (runs a process that a box would serve), integrator (installs systems for many sites), software vendor (whose product would ship on a box). Distributors and publishers are kept but not counted as prospects.

How fit is scored

Claude Haiku reads each company's evidence and your offer and returns a score from 0 to 100 with two to four reasons. The code then checks the answer:

Competitors

For each offer, SiliconMatch also asks whether a company sells something that directly competes with it, and keeps the flag only with that company's own evidence. The level matters: if you sell chips, only other chip vendors compete (module, board and drone makers buy chips and stay prospects); if you sell a box, companies selling the same type of chip or their own edge-AI hardware compete, while software that would run on your box counts as a partner. Competitors are hidden from results until you tick Show competitors, and then appear last with a badge and the evidence.

Signals, map, trends, graph

Signals. Tech, AI, semiconductor, robotics, drone, energy and EV news feeds, Google News searches per prospect, verified job boards (Greenhouse, Ashby, Lever) and public tenders (EU TED; US federal awards on USAspending.gov). An item becomes a company's signal only when it names the company: multi-word names as a phrase, single-word names only next to a company cue (“Augury raises…”), never a lone common word. Job posts count only when the title names hardware skills; tenders only when they are about technology.

Why now. A move in the last 90 days (funding, expansion, tender, partnership, product, a hardware job post) adds to the fit, more when it names your offer's own terms, and never more than 5 points. The stored score is not changed, and the line always cites the signal's date, headline, source and link.

Sites. Power plants from the WRI Global Power Plant Database, hospitals from OpenStreetMap and US factories from the EPA greenhouse-gas reporting program, linked to a prospect only when the registry's owner, operator or parent name matches the company by our name rules.

Forecasts. Each series is backtested on rolling cut-offs; a forecast is shown only if it beats a same-month-last-year forecast and the series is not mostly zeros. The model is Amazon Chronos-Bolt (Apache-2.0); bands are its 10 % and 90 % quantiles.

Graph. Links are built from quoted sources only. The words around a name decide the link (“built on NVIDIA Jetson” = uses; “partnerships with” = partners with); names that only appear in the same list are “listed with”, peers rather than partners. Links the sources do not state are not drawn.

Papers. Recent research papers come from OpenAlex, for companies whose OpenAlex record matches their website or name. US patents are not included because the patent office's API now requires identity verification.

Practice room

Each simulated buyer is built from one real prospect's dossier: a card of numbered facts from quoted sources only (no AI-written summaries). It answers as that company's engineering or operations lead, citing its facts for anything company-specific and saying “typical for the segment” otherwise; answers citing nothing from their own card are marked uncited. Competitor agents come from the companies flagged as competitors. In round two each buyer sees its segment's round-one split (anonymous numbers), which can create herding; the page reports how many changed their mind. All splits are counted by the code. Results are hypotheses: once 20 real outcomes are marked on your lists (replied, meeting, not a fit), the page reports how often the simulation ranked them the right way.

Privacy and limits

Architecture

A Python web service (FastAPI) with a SQLite database, a plain HTML and JavaScript front end, and outside services: Claude through OpenRouter (Haiku for search, extraction, scoring, drafting, voice intent, briefs and simulations; Sonnet as a fallback), ElevenLabs (speech to text and Ellen's voice) and a contact provider (Hunter). A nightly job refreshes news, job boards, tenders, papers, the graph and each seller's brief; forecasts run on a small local model. The demo runs in two lanes with the same code: on our own Mac (chipapp.wntap.org, reached through a secure tunnel) and as a container on a Hetzner cloud server (siliconmatchapp.dkube.app, deployed from its repository on every change, database backed up daily). Each lane keeps its own data and refreshes it nightly. This page is served at both chip.wntap.org and siliconmatch.dkube.app; its demo links follow the address you came in on.

browser → chipapp.wntap.org (Vercel front) → tunnel → path router → SiliconMatch on the Mac
browser → siliconmatchapp.dkube.app (Vercel front) → Hetzner (nginx, TLS) → SiliconMatch container
                                                        ├─ SQLite: companies · evidence · scores · contacts · lists
                                                        ├─ Claude Haiku (OpenRouter): search, extract, score, draft
                                                        ├─ Hunter: contacts (not from the public demo)
                                                        ├─ ElevenLabs: speech to text, Ellen's voice
                                                        └─ nightly: news · job boards · TED / USAspending · OpenAlex · graph · briefs

API reference

All endpoints except sign-in need a signed-in session (a cookie set by /api/login).

Method and pathWhat it does
POST /api/loginSign in with {"pin": "1111"}
GET /api/capabilities, /api/segments, /api/boxThe dropdown, the segment groups and the box parts
POST /api/searchStart a search: {capability, box[], segments[], country}; returns a job id
GET /api/search/{job}Progress: stage, percent, candidates, scored
GET /api/results?capability=&box=&segments=&country=Scored prospects with reasons, evidence, opener, roles, integration
GET /api/companies/{id}/peopleStored contacts, ranked by role
POST /api/contacts/{id}/sequenceDraft a three-step sequence for one contact
PUT /api/steps/{id}Save an edited step
GET/POST /api/lists, GET /api/lists/{id}Your lists; create one with {name, offer}
POST/PATCH/DELETE /api/lists/{id}/items/{company}Save, update status or note, remove
GET /api/lists/{id}/export.csvDownload the list
GET /api/dossier/{domain}, /dossier/{domain}.pdfEverything known about one company; the printable brief
GET /api/map/features?capability=&box=Pins, country counts and sites for the map
GET/POST /api/watchlists, POST /api/watch/{company}, GET /api/inboxWatchlists and the inbox digest
GET /api/series, GET /api/heatingMonthly series with forecasts; markets heating up
GET /api/graph/questions, /api/graph/ask/{id}, /api/graph/ego?node=Path questions and a node's neighbourhood, with evidence
POST /api/voice (audio), POST /api/voice/textA spoken or typed command: transcript, chosen action, where to go
GET /api/brief, /api/brief/audioYour latest morning brief and its audio
POST /api/sim, GET /api/sim/job/{id}, /api/sim/runs/{id}, /api/sim/calibrationRun a practice-room scenario; its result; the calibration report

Glossary