How we built pre-raise founder discovery for an onchain fundraising platform
An onchain private fundraising platform wanted to reach strong founders before they start a formal round. Releasi built a recurring discovery system that watches accelerator cohorts, founder posts, hiring and specialist news, links each company to its founders and keeps the dated evidence on every record. By 29 October 2026 it held 1,937 active companies and 2,829 linked founders.
The brief
The client runs a platform for compliant private fundraising on onchain rails. It wanted conversations with strong founders early, whether or not they had started a formal round, plus a clear view of who to speak to and why.
Two rules shaped the build. A credible source only makes a company a candidate, so discovery and qualification stay separate. And every record carries the dated public evidence behind it.
What the system watches
| Lane | What it looks for | Runs |
|---|---|---|
| Accelerator cohorts | Published YC 2025 and 2026 batches, current a16z Speedrun and Alliance cohorts | Every 48 hours |
| Application disclosures | Founders posting on X that they applied to or joined YC, Speedrun, Alliance, Antler, Entrepreneur First, South Park Commons, HF0 or PearX | Every 48 hours |
| Founding hires | Posts recruiting a founding engineer, marketer or GTM lead | Every 48 hours |
| Commercial hiring | New sales, growth and GTM roles at known startups | Daily |
| Launches | Show HN and App Store releases | Every 48 hours |
| Specialist news | Dated startup milestones in deep-tech trade press | Weekly |
| Early departures | First-person posts about leaving a job to build something new | Weekly pilot |
How a signal becomes a record
- Collect. Each source has its own collector, schedule and saved evidence, so a broken feed can be spotted and fixed alone.
- Screen. Cheap rules run first. A small language model then picks the relevant passage, and code copies the real text, holds invented names and never guesses a website.
- Resolve the people. Founders and LinkedIn profiles come from the programmes' own pages. Personal X accounts stay as candidates until something corroborates them, and a matching name never merges two people.
- Review. Discovery, qualification and contact readiness are separate fields. Human decisions survive every refresh, and rejected companies keep their reasons in an archive.
Where it stood on 29 October 2026
- 1,937 active companies, with 1,516 in the priority tier and 421 in a second tier. Most of this starting inventory comes from published YC batches.
- 2,829 founders linked to those companies, with 2,798 LinkedIn profiles and 1,541 personal X profiles recovered from programme pages.
- 56 companies archived as poor fits, with the reasons kept.
What it has not proven yet
- Outreach was not connected at this date, so there are no reply or meeting numbers to report.
- No emails were bought or guessed, so email coverage is still unknown.
- The first early-departure run screened 33 X posts and found no valid announcement. That lane is still a pilot.
- Crypto-specific feeds such as ETHGlobal projects and DefiLlama protocols are ideas for this build and are not connected.
The stack
Custom Python collectors and a SQLite store on one server, with a review dashboard hosted on Vercel. TwitterAPI.io supplies X search, small OpenAI models handle structured screening, and Apollo runs bounded company searches. There are no Clay tables in this build.
Commercial platforms such as Harmonic, Evertrace and Frontrun cover far larger populations. This build is narrower and follows one team's definition of a good founder.
Who else this fits
Any company selling to founders near a raise: fundraising platforms, launchpads, market makers, token counsel and cap table software. The lanes change with the buyer, and the review model stays the same.
Nicolas Goehler is the founder of Releasi, which finds B2B companies showing public signs of the problem you solve. Read the intent signals playbook or see his code on GitHub.