My co-founder Prapti and I just opened the waitlist for Something, a founder/investor matching platform — but I wanted to write for this audience specifically about the technical part, since that's usually more interesting here than the pitch.
The core problem with most AI feedback tools: they're sycophantic by default. Optimized for engagement, which nudges toward validation. We wanted the opposite for the moment someone's deciding whether a startup idea is worth building.
So before any idea on the platform goes public, it runs through an adversarial multi-agent pipeline:
- One agent argues the strongest possible case for the idea
- A second agent — we call it Nothing — is specifically rewarded for surfacing flaws: weak unit economics, bad market timing, technical infeasibility
- Output is a structured critique + conviction score, not a vague "looks promising!"
Some of the architecture decisions behind it:
- Multi-agent orchestration (LangGraph) instead of a single-prompt approach — lets the optimist and skeptic reason independently before reconciling
- Hybrid retrieval (dense + sparse) to ground critiques in real market data instead of hallucinated reasoning
- Cost-aware model routing across providers — routes to cheaper/faster models where the task doesn't need frontier reasoning, keeps inference costs sane at scale
Once an idea survives review and the founder has actual proof of work (repo, pilot, patent, live demo — no slide decks), it gets matched to investors based on their real deployment history instead of cold outreach.
Waitlist's open if you want to poke at it: https://something-waitlist.vercel.app
Happy to go deeper on any part of the architecture in the comments — genuinely want pushback if something sounds off.
Top comments (0)