Method
A three-person company building a reactor or an interceptor has no headcount or web traffic to track. It still has to file with the SEC to raise, the FCC to test a radio, the FAA to fly and the NRC to discuss a reactor. Antenna reads those filings, matches them to code, papers and job posts, and ranks higher the companies that appear in more than one kind of source.
Pipeline
- Collect4,729 signals from eight families
- Match3,082 companies
- Score826 on thesis
- Rank300 ranked
The score
Edge = momentum × thesis fit × earliness × team
Momentum
Each signal has a strength from 0 to 100 that halves on a schedule set by its family. A second signal in the same family adds less than the first. A company with signals in only one family is capped and marked down, so a signal from a second family is what moves it up.
Thesis fit
Everything written about the company (its own site, award abstracts, filings, job posts) is matched against term lists for robotics, autonomy, defense, energy, manufacturing, semiconductors and space. It is a keyword match, and the dossier lists the terms that matched.
Earliness
How little the company is already known. GitHub stars, Hacker News mentions, traffic rank, open roles, money raised, age and a later reviewed stage each lower it, but only above a floor: $15M raised, 10 open roles, 500 stars, three years old. A company with nothing to measure is marked not measured and is not scored as unknown.
Team
Affiliations are quoted from the source and matched against a table of forty labs and companies that hard-tech founders tend to come from. Team lifts a score a little and never sinks it, because most early filings name no one but the signer.
Families
Sources
Sanity check
Portfolio companies are kept off the board. The pipeline found these anyway.
- Saronic Technologies
Review
471 companies reviewed, 132 removed. Nothing is ranked until it has been reviewed.
Each company is checked by an agent told to assume the row is wrong. It opens the evidence, looks the company up and reports what it is: formation, early, growth, established or not a company. The stage on each row comes from that.
- An FAA waiver is held by whoever flies. Survey operators and dealers were being read as drone makers.
- One person can open an NRC docket with a letter, so a docket alone proves little.
- The people on a filing are whoever signed it: a regulatory lead, outside counsel, a sales engineer.
- Subsidiaries of public companies file under fresh legal names and look like startups.
493 companies that fit the thesis are waiting for review and are not on the board yet.
Where it is weak
Paperwork is filed by old companies too.
A first trademark or a new FCC test authority looks the same from a company formed last month and from Robin Radar, sixteen years old, which filed for a sales demo and ranked in the thirties. Nothing in a filing says how old or how funded the filer is.
First means first in one registry.
A first federal award is first under one identifier, a first Form D is first under one SEC number, and a renewed waiver reads like a new one. None of them means the company is new.
Rank history is reconstructed.
These are the first runs. Past weeks are recomputed from event dates, which is accurate for filings and awards and approximate for anything measured only today, such as DNS records.
Defense awards arrive late.
Department of Defense procurement data is embargoed for ninety days, so an Other Transaction on the board was signed at least three months ago. Form D and FCC filings are the timely ones.
GitHub says little about hardware.
GitHub is rich for robotics, simulation and chip design and nearly silent for nuclear, grid and space. A popular repository is also often one person, not a company.
Hiring pace cannot be measured yet.
Job boards show only the roles open now, so filled roles vanish and a bulk repost looks like a hiring spree. Measuring the pace needs a second daily run.
Legal names are not brand names.
A filing says Aalo Holdings and the press says Aalo Atomics. A match on the legal name can miss, the same company can appear twice, and a Hacker News search under the legal name finds nothing. So an empty search is never scored as a sign the company is unknown.
No paid data, no private data.
Everything comes from free public sources. There is no LinkedIn, X or PitchBook data.
The weights are a guess.
Family weights were set by hand. They have not been tested against outcomes.
Lead time
2 of 24 briefed companies had a signal on file before they were first mentioned anywhere in public.
Any earlier mention counts as public: a launch post, a line in a local paper, a name on a grant list. Lead over the announcement of a round would be the more useful measure and cannot be tracked until there are daily runs to compare.
What comes next
- 01
Funding and founding data
Well-funded companies can look new because the scorer does not know what they have raised. A licensed funding source would replace most of the review step.
- 02
Partner decisions
Record each meet or pass and use it to set the weights.
- 03
Daily snapshots
Store every daily run. Hiring pace, DNS changes and rank history can then be measured instead of reconstructed.
- 04
Engineers leaving together
Several engineers leaving the same team at SpaceX or Anduril in one quarter often means a new company. This needs paid employment data.
- 05
Briefs for the whole board
Briefs exist for 24 companies. Each takes an agent a few minutes, so they could be written nightly for every company that moved.