Stringers
ChazY11/medradar · ChazY11

9 AI agents built an FDA radar in 11 hours, and 24 stars that arrive at a steady 159 a day

What it is A free, open-source Chinese tool: it uses public data like FDA filings, clinical trials and US-listed company announcements to track where the big US medical device and drug makers are shifting, then produces research reports, dashboards and a paper-trading portfolio.

Why it took off It didn't really take off: three days after launch it had 0 stars, and since then ten to twenty a day have trickled in. Most of the accounts that starred it have 0 followers and were registered in batches on the same day, and they share a pool with a group of Chinese AI repos launched around the same time. Nobody on X, HN or in the industry circles has mentioned it, it has only 9 forks, and every issue and PR from outsiders is 0.

DHow you could play it Don't copy it. If you really want to build something, go after the half it gave up on: China's NMPA registrations and volume-based procurement wins. Or stick to one sector and update it weekly, then launch it with verifiable numbers in the framework vendor's showcase and to industry readers.

D grade: the stars were manufactured, so they prove nothing about whether anyone wants this kind of project. The field is cold on GitHub, and the real buyers aren't on GitHub. What's worth learning is topic selection and presentation, not copying it. Almost nobody knows about it yet
Data (updated daily) · Repo created 2026-10-31 · 308 stars · 7 days +127
GitHub · stars
010020030011-0511-0911-1311-1611-20308
Data 2026-11-05 → 2026-11-20 (16 days, 16 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

MedRadar is a Chinese-language "US FDA industry direction radar": it turns free public data from the FDA (device approvals, recalls, clinical trials) and US stock announcements into sector heat rankings, big-company movement tracking, event studies, and an offline dashboard. The author says its core was built in one day by 9 AI agents. It was built 12 days ago and has 159 stars. But those stars didn't come from people in medicine or investing; they came in steady batches from empty accounts registered in bulk. The product is very full and worth a look, but this curve teaches nothing about how to actually get people to star a repo.

One day in Oct 31, and the first star in Nov 3

On 2026 31, GitHub user ChazY11 created the medradar repo. According to its own attribution doc, the work that day was done by a team set up with DeepSeek Harness (DeepSeek's open-source agent orchestration framework): 1 Leads handling integration and acceptance, and 8 specialist agents covering device data, drug data, quant, visualization, the product dictionary, replay, stats and packaging. The model was deepseek-flash, with "about 11 hours of continuous autonomous iteration." The doc even has a line on this:

External models: none (no non-Chinese models were used)

A blue badge at the top of the README reads "Built by DeepSeek Harness · deepseek-flash."

For the next three days, the repo had zero stars.

On Nov 3 20:03 (UTC), the first star arrived from william-stamper242o, with 0 followers, registered on 2025-04-15 12:13. The fifth star that day came from samuel-edwards487y, also with 0 followers, registered the same day, 11:58, 15 minutes before the previous one, and its bio reads "Frontend architect | Angular specialist." The sixth star was the author's own.

From that day on, 9 to 24 a day, and someone starred it at every hour of the day, roughly as often at 2 a.m. as at 2 p.m. (from GitHub's public event stream, pulled 24 11-12).

What it is

A Python command-line tool, plus a 3.3 MB HTML dashboard that doesn't connect to the internet (on GitHub Pages: chazy11.github.io/medradar). The question it sets out to answer is stated in the README: "Don't guess what big companies 'want to do'; count what they've already done and where they're moving."

  • Data: openFDA's 510(k), PMA and recalls; ClinicalTrials.gov; SEC 8-K announcements; Yahoo quotes, all free. A full run ingests 375k events into a 1.2 GB SQLite database.
  • Output: heat rankings for 60 tracks (in 10-31, the hottest were cardiac rhythm, wearable monitoring, structural heart disease, and surgical robotics), product-mix shifts at 89 major companies, a list of "first-ever product codes" for those companies, 204 device product families and 2,740 Chinese product names, and 19 chapters of Chinese research reports.
  • On the stock layer: run an event study on FDA events to see whether they move the share price, then plug in a long-only paper portfolio where an LLM reads the full 8-K text and decides whether it's good or bad news and how big the position should be.
  • Try it: medradar demo, offline, no API key needed, results in 60 seconds.

It is candid in its writing: where something can't be done, it says so. For example, China's NMPA data is marked "unavailable," and when FDA advisory committee calendar scraping fails it reports empty rather than pretending it worked. Of the repo's 30 PRs, all were opened by the author, and 5 of them have titles like "NO-GO verdict" and "negative result." These are cases where it had its own agents test whether FDA signals could foresee big companies' R&D direction, and the answer was mostly that they couldn't. The README also says the newly added data sources have "no statistically detectable ranking gain."

Where these 159 stars came from

Look at the curve first. Daily star snapshots from GitHub: 11-05 28, 11-06 37,11-07 52,11-08 64,11-09 87,11-10 109,11-11 131,11-12 154, later reaching 159. The daily gains are roughly steady, with no single day spiking from someone's repost.

Next, look at who gave the stars. Of the 160 accounts that starred it:

What to look atResult
0 followers124
Followers no more than 1142
Registration dates cluster2025-04-15 registered 10, with bios all reading 「Mobile developer \Flutter & Dart」「Web developer \Templates like "Vue.js specialist"; 2026-05-10 and 2025-08-09 each with 8 entries
A run of accounts with names starting with hawaii and aloha, and a run whose bios say "[some city] No. 1 Middle School"A batch of accounts starting with hawaii and aloha, and another batch with bios reading "Some City No. 1 High School"

We sampled 12 random accounts and checked what else they had starred: 12 of them starred both agnes-harness and erupt, 11 starred Deepgraph, and 10 starred moworld_base and LiveStream-Agent-Studio. These repos were all climbing the same week, and several of them had over a hundred stars with forks at 0. Another repo billing itself as a "DeepSeek Harness research tool," DKKD13/dsh-finance, was created 11-01, has 126 stars and 0 forks, and 16 of its starrers also starred medradar.

Last check, is anyone actually using it: 9 forks, 0 issues from outsiders, 0 PRs from outsiders. Nothing on X in nearly 30 days, nothing on Hacker News, nothing in the GitHub discussions, nothing in DeepSeek Harness's official showcase either. A web search only turns up the repo itself. The author account has 10 followers, and the author's other 3 repos each have 0 stars.

So the conclusion is fairly clear: these stars came from a pool of accounts that star-farm or trade stars with each other. Whether the author bought them or joined a mutual-star group, we can't tell from the outside. A few accounts look like real people (one bio even lists a Harvard biology PhD), but that's a very small share.

What it got right

Setting aside where the stars came from, a few things about this project are worth noting.

We picked a GitHub niche that's almost empty. Searching the fda, openfda, and medical-devices tags, its 154 stars rank 7, with only the official FDA repo (723 stars) and a few NVIDIA medical projects ahead. On the open-source side, FDA projects are either raw official APIs or query tools for LLMs (medical-mcp at 121 stars, us-gov-open-data-mcp at 112 stars). This is the only one that turns it into a finished research report plus dashboard.

This space is expensive to play in. The ones doing the same thing are paid databases:

CompetitorsWhat it isHow it differs
Paid regulatory intelligence for device companiesEnglish, expensive, human-analyzed; no stock event studiesEnglish-language, pricey, with human analysis; no stock event studies
One of China's leading drug and device intelligence platformsChinese-language, with domestic registration and procurement data; MedRadar covers only US dataBioPharmCatalyst, BiopharmaWatch
BioPharmCatalyst, BiopharmaWatchBiotech catalyst calendar for retail traders, about $29/monthFree, open-source, Chinese-language, reproducible offline: that difference is real. But the buyers are device companies and medical investors, and they use the databases above. They don't come to GitHub to look for tools, and the author hasn't gone to where they are to say anything.

Free, open-source, Chinese-language, and reproducible offline: that difference is real. But the buyers are device companies and medtech investors, and they use the databases above. They won't go looking for tools on GitHub, and the author hasn't gone anywhere they hang out to say a word about it.

The stories and numbers are ready-made demo material. "9 agents, 11 hours", "375k events", "678 offline tests", "no non-domestic models used". This is exactly the kind of case model and framework teams love to share. DeepSeek Harness went open source in August and now has 279k stars, gaining over six thousand in a single week, and the community is short on good showcase projects. It just hasn't been picked up yet.

If it were you

Who it suits: People already working in devices, pharma or investing, with an audience of their own (WeChat Official Accounts, research chat groups, investor circles), who want to build an open-source tool to show peers. For you, it's a ready-made list of topics and data sources.

Who it's for: Anyone already in medtech, pharma, or investing, with an existing audience (WeChat Official Account, research chat groups, investor circles) who wants to build an open-source tool their peers would actually use. For you, it's a ready-made list of topics and data sources.

How to do it: If you still want to build in this space, pick the part it did badly. One is the China half: it marks domestic NMPA data as "unavailable," yet domestic registrations and volume-based procurement wins are exactly what Chinese device companies and investors most want to see, and where it falls furthest short of MedDevice NextDevice. Two is narrowing: don't cover 150 sectors, pick one (say surgical robots or brain-computer interfaces) and update it weekly, so people in the industry want to share it.

How to get stars: Launch is exactly what it lacked. After you build, post to the official showcase of the model or framework vendor, or to their Discussions, with verifiable numbers and a demo that runs in 60 seconds, aiming for one official repost. At the same time, post an article for industry readers titled something like "What the FDA approved this week," and put the repo link at the end. Relying only on organic traffic from GitHub's trending tabs won't bring people to a cold niche.

How to get stars: What it's missing is launch exposure. After you build it, submit to the official showcase or Discussions of the model or framework maintainers, with verifiable numbers and a demo that runs in 60 seconds, aiming for one official repost. Then post a "what the FDA approved this week" piece to industry readers and put the repo link at the end. In a niche field, waiting on natural GitHub traffic alone won't get you anywhere.

What would show you're wrong: If the stars keep climbing after the star-farming stops, or outsiders start opening issues and people in the medical or investing circles begin sharing it, that means the field really does have an audience on GitHub waiting for it, and it's worth serious work. If DeepSeek Harness officially reposts it and it brings a wave of real stars, that means the route of serving as a showcase case for a framework vendor can work.

Signs you're reading it wrong: If stars keep climbing after the spam stops, or outside users start filing issues, or people in medical or investing circles start sharing it, then the field has a real audience on GitHub and is worth building seriously. If the DeepSeek Harness team reposts it and sends real stars, then pitching yourself as a showcase for framework makers can work.

GitHub repo ↗

Research notes: competitor table, acquisition evidence, full research report
Competitors
NameWhereLaunchedScaleHow it differs from this one
FDA/openfdaOpen source2014723 starPaid regulatory intelligence SaaS for device companies
Basil Systems MedTech IntelOtherYears agoPaid, English-language, searchable across PMA/510(k)/recalls/adverse events; MedRadar is free and open-source and adds stock event studiesPaid, English, searchable 510(k)/PMA/recall/adverse event data; medradar is free and open source, and also adds stock-price event studies
Industry-standard paid databasesOtherYears agoInclude human analysis and market forecasts, and are expensive; MedRadar is free and fully automatedHuman analysts and market forecasts cost a lot; medradar is free and fully automated
Medical Radar NextDevice / NextPharmaWebYears agoChinese-language, with domestic registration and procurement data; MedRadar covers only US dataChinese language, covers NMPA and tender data; medradar only does public US data
BioPharmCatalyst / BiopharmaWatch / bpiqWebYears agoA biotech catalyst calendar for retail investors, about $29/monthLeans toward drug PDUFA dates and trading; medradar leans toward device industry and sector heat
JamesANZ/medical-mcp, lzinga/us-gov-open-data-mcpOpen source2025-07 / 2026-02121 / 112 starAn FDA query tool for LLMs, not a finished report
Where users come from
  • Confirmed Star-farming and mutual-star account pool: 160 WatchEvents. The first star arrives at 11-03 20:03 UTC, then 9~24 a day, with stars in all 24 hours. 124/160 accounts have 0 followers. Registration dates cluster (2025-04-15 has 10 accounts with templated bios; 2026-05-10 and 2025-08-09 have 8 each). Of 12 sampled accounts, 12/12 starred agnes-harness and erupt, and 11/12 starred Deepgraph. This batch of repos rose at the same time.
  • Inferred DeepSeek Harness attribution and the domestic-model narrative: the README header carries a Built with DeepSeek Harness badge, and docs/ATTRIBUTION.md says 1 Lead + 8 agents built it in 11 hours, "no non-domestic models." But no DSH Discussions, awesome lists, or X mentions turn up.
  • Confirmed GitHub topics and the Pages dashboard: 12 topics, and chazy11.github.io/medradar is a self-contained dashboard. Among FDA-related repos it ranks #7 by stars.

Why ChazY11/medradar could be built

Verdict

MedRadar is a Chinese-language "US FDA industry opportunity radar." It uses free public data: openFDA (510(k)/PMA/recalls), ClinicalTrials.gov, SEC EDGAR 8-K, and Yahoo quotes. It produces sector heat, big-company momentum, first-seen products, event studies, and paper-trading replays, outputs a 19-chapter report, and includes an offline HTML dashboard (GitHub Pages: chazy11.github.io/medradar). The README's first line has a "Built with DeepSeek Harness · deepseek-flash" badge, and docs/ATTRIBUTION.md says the core was built by 1 Leads and 8 agents running deepseek-flash for about 11 hours straight.

Most of these stars didn't come from the product itself. They came from a star-farming network. The evidence:

  • For the first 3 days after creation, the repo had zero stars. The first star appeared at 11-03 20:03 UTC, then 9 to 24 a day, with stars in all 24 hours. The spread is even, a classic drip. The author also starred it on 11-03 22:58.
  • Of the 160 accounts that starred it, 124 have 0 followers, and 142 have no more than 1. Registration dates come in batches: 10 accounts were created on 2025-04-15, with bios like "Mobile developer | Flutter & Dart" or "Web developer | Vue.js specialist"; 8 were created on 2026-05-10; 8 on 2025-08-09. There's also a run of hawaii*/aloha* accounts, and a run whose bios read "No. 1 High School of [a city]."
  • I sampled 12 accounts to see what else they starred. 12 starred all of AgnesAI-Labs/agnes-harness and erupts/erupt, 11 starred billion-token-one-task/Deepgraph, and 10 starred Moxin-Tech/moworld_base and HanyuanWang/LiveStream-Agent-Studio. This batch of repos rose at the same time.
  • No external spread at all: 30 posts on X (x-radar, last 0 days), 0 on HN, 0 GitHub Discussions threads, and nothing in DSH's official "Show Your Plugins" board. Jina search returns only the repo itself.
  • Of the 154 stars, only 8 forks and 0 issues. The outside PRs are 0, and all 30 PRs were merged by ChazY11 himself.

The product itself is genuinely thorough. 375k events, a 1.2 GiB SQLite database, 678 offline tests, a 19-chapter report, a self-contained dashboard, and a "60-second offline demo." The README also makes many "honest" statements (unusable things are labeled unusable, negative results are kept). So it can look worth starring, but looking worth it and people actually starring it are different things. Very few real people starred it.

1. Competitors: who does the same thing

NameWhereWhen it launchedScaleHow it differs from medradar
FDA/openfda (official)Open source / official API2014723 stars (GitHub repo, 11-12)Raw data interfaces; medradar aggregates, scores, and writes Chinese reports on top of them
Basil Systems MedTech IntelCommercial SaaSYearsPaid regulatory intelligence for device companies (basilsystems.com)Searches 510(k), PMA, product codes, recalls, and adverse events. Paid, English-only, no stock-event studies
Industry-standard paid databasesCommercialYearsInclude human analysis and market forecasts, and are expensive; MedRadar is free and fully automatedHas human analysis and market-size forecasts, and is expensive. medradar is free and fully automated
Medical Radar NextDevice / NextPharmaDomestic commercialYearsOne of China's leading drug and device intelligence providers (pharmcube.com)Chinese language, covers NMPA and tender data. medradar only does public US data, and marks NMPA as honestly unavailable
BioPharmCatalyst, BiopharmaWatch, bpiqWebYearsBiotech catalyst calendars for retail investors, with a free tier and paid plans around $29/monthLeans toward drug PDUFA dates and trading. medradar leans toward device industry
JamesANZ/medical-mcp, lzinga/us-gov-open-data-mcpOpen source2025-07 / 2026-02121 / 112 starFDA query tools for LLMs. medradar is a complete finished report
innolitics/rdm, rOpenHealth/openfda, coderxio/sagerxOpen source2015–2021144 / 65 / 82 starRegulatory docs or data pipelines, not industry radars

Searching GitHub by topic (fda/openfda/medical-devices) or keywords, medradar's 154 stars already rank #7. Above it are only the official FDA repos and a few NVIDIA medical repos. The field is cold on GitHub to begin with, and no open-source FDA industry radar exists yet. There's no direct competitor, but the people who need this (device companies, investors) use paid databases and won't come looking on GitHub.

II. Growth levers: where users come from

  1. Star-farming and mutual-star networks (the main source, largely confirmed). Evidence:
    • Star timeline (160 WatchEvents from the GitHub events API, covering all stars): 11-03 9, 11-04 19,11-05 11,11-06 17,11-07 13,11-08 24,11-09 20,11-10 23,11-11 22. Stars arrive in every hour, with no single burst window.
    • Account profiles are above. Of the 12 sampled accounts, the repos they co-starred include agnes-harness (12/12), erupt (12/12), Deepgraph (11/12), moworld_base (10/12), LiveStream-Agent-Studio (10/12), OPEI (9/12), lazy-balancer-v2, pulseaiclub/phi, and nuglifeleoji/Sentry (8/12 each). On GitHub these repos have all risen over the last 7 days: Deepgraph +1308, Sentry +678, agnes-harness +340, fluxdown +134 with 0 forks, and OPEI +121 with 0 forks.
    • The same pool contains another "DSH-built research tool": DKKD13/dsh-finance (created 11-01, 126 stars, 0 forks). Of its 125 starrers, 16 also starred medradar.
  2. "Built with DeepSeek Harness" badge to ride the hype (present, but no visible effect). DSH 2026-08-13 is open source, now at 279k stars, up 6532 in the last 7 days (GitHub repo). medradar puts the badge at the top of its README and writes a whole page on the credit, stressing "no non-Chinese models used," clearly aiming to tie into DSH and the domestic-model conversation. But we couldn't find any sign the DSH community actually picked it up (Discussions, awesome lists and X searches all come up empty).
  3. GitHub Pages dashboard + topics. The 12 topics (fda, openfda, sec-edgar, and so on) get it onto topic pages, but organic traffic from those is tiny.
  4. None: no mentions on HN, X, Reddit, Xiaohongshu, Zhihu, or in the media. The author has 10 followers, no X account, and no blog. The author's other 3 repos each have 0 stars (a Codex desktop pet, AR).

3. Why it ran up numbers (or rather, why the data makes it look like it did)

  • The stars come from a star-farming or mutual-starring account pool. The drip has been even since 11-03. Most accounts have 0 followers, were registered in bulk, and have templated bios. They share the same pool with a batch of other new Chinese AI repos from the same period.
  • A story of "AI agents built an over-spec data product in 11 hours," plus a set of numbers to show off (375k events, 678 tests, 19 report chapters), makes the farmed stars look reasonable. It's also a good showcase for DSH or domestic models.
  • The topic avoids crowded ground. FDA device intelligence is nearly empty on GitHub, so 150 stars is enough to reach the top of a topic list.
  • The author commits constantly (184 commits on main, 30+ research/* branches, and still pushing 11-11). The repo looks very active, but outside participation is 0.

4. Differentiation

Compared with paid databases (Basil, Evaluate, Pharmcube): free, open source, Chinese-language, and reproducible offline, and it chains FDA events with stock-event studies and paper trading. Compared with open-source peers: other open projects are just openFDA clients or MCP servers, while this one delivers a finished report and dashboard.

The differentiation is real, but it is not why it got its stars. The stars were farmed, and nobody in medical or investment circles is discussing it. The target users (Chinese device companies, medical investors) don't look for tools on GitHub, and the author hasn't promoted it where they are.

5. Visibility and copycats

  • Visibility: still very low. 0 posts on X, 0 on HN, no media coverage, no Discussions, and external issues and PRs are 0.
  • Copycats: none that copy it directly. DKKD13/dsh-finance in the same star-farming pool uses the same playbook (DSH-built, research tool, evidence library, paper fund). That's a similar move, not a copy.

6. What to borrow

  • Pick a data domain that is cold on GitHub but commercially expensive (paid databases often cost tens of thousands a year), and use free public data plus batch agent engineering to build a "cheaper alternative radar." 150 stars is enough to reach the top of a topic list.
  • "Agent team builds it in N hours" plus a checkable list of numbers plus an attribution page makes good material for model or framework vendors to showcase, and could win official reposts (which it hasn't gotten).
  • Put a self-contained HTML dashboard on GitHub Pages and a 60-second offline demo to lower the barrier to trying it.
  • Lesson: star-farming alone doesn't bring forks, issues, or real users, and the real users in this field aren't on GitHub.

VII. Uncertain

  • The stargazers endpoint returned 404 for our token. Stargazer accounts came from the events API WatchEvents (160 entries, roughly matching the 159 stars), so I couldn't verify individually whether all of them are farm accounts. A few accounts look like real people (KeikaJames has 16 followers, and tsneala's bio says Harvard biology PhD), but that's a small share.
  • Not sure whether the author bought stars or joined a mutual-star group (the author also starred it, and I didn't dig into the repos they starred).
  • Don't know whether DSH officials or domestic communities (WeChat, QQ) shared it. Those channels aren't searchable, so I can't rule it out.

Sources

  • data card /home/ljr/fanout/why/cards/gh_ChazY11__medradar.json (GitHub repo, 11-05~11-12 daily stars)
  • GitHub API (2026-11-12 pulls): repos/ChazY11/medradar, /events (WatchEvent, ForkEvent), /forks, /pulls, /branches, users/<starrer account>, users/<account>/starred
  • README.md, README.zh-CN.md, docs/ATTRIBUTION.md (github.com/ChazY11/medradar)
  • GitHub repo_profile: FDA-category repo rankings; star counts and gain_7d for agnes-harness, Deepgraph, Sentry, dsh-finance, deepseek-ai/deepseek-harness
  • x-radar search for "medradar", "ChazY11", "chazy11.github.io" over the last 30 days, 0 results; hackernews library 0 stories; GitHub Discussions search 0 results
  • Jina search: deepseek.com/en/harness, github.com/deepseek-ai/deepseek-harness/discussions, basilsystems.com, evaluate.com, pharmcube.com, biopharmcatalyst.com, biopharmawatch.com

This article was written by AI from public data and the data we collect daily. Numbers are sourced, and inferences are labeled.