Thesis Killer

Note for judges: This repository is private. Access has been shared with mlempowerfdn@gmail.com (GitHub collaborator invite pending). Contact us if you need alternate access.

Inspiration

Every investor falls in love with their own thesis — and nobody wants to argue against themselves. We built Thesis Killer to do that job automatically: take any investment idea and try to prove it wrong, using real, current data instead of gut feeling.

What it does

You type your investment thesis in one sentence. Thesis Killer scans it across four angles — valuation, competition, macro conditions, and execution risk — searching the live web for evidence that contradicts or weakens it. It returns a fragility score (0–100), a verdict (SOLID / DEBATABLE / FRAGILE / DEAD), the specific cracks it found with real data behind each one, the single best argument for the thesis (so it stays honest, not just negative), and the one measurable event that would actually kill the thesis.

How we built it

The frontend is a React interface built for real-time feedback — a live "scanning" sequence across the four risk pillars, an animated fragility gauge, and a stamped verdict. It calls the Claude API directly with web search enabled, so every autopsy runs against fresh data instead of stale training knowledge. The system prompt forces a strict JSON output (score, verdict, flaws, steelman, kill condition), which the frontend parses and renders live.

Challenges we ran into

Getting an LLM to be a fair judge instead of either a cheerleader or a doom machine took real prompt tuning — we had to explicitly force it to output the steelman argument for the thesis, not just attack it, so the score reflects real evidence rather than a fixed negative bias. Parsing reliable structured JSON out of a model that also does live web search required defensive extraction on the client side.

Accomplishments that we're proud of

A working, honest "adversarial" analysis tool that reacts differently to different theses — tested side by side on a fragile thesis and a solid one, and got two genuinely different, evidence-based verdicts instead of a canned response.

What we learned

The hardest part of building a useful financial AI tool isn't the interface — it's designing the incentive structure of the prompt so the model argues honestly instead of pleasing the user or performing pessimism.

What's next for Thesis Killer

A configurable time horizon (does this thesis hold at 3 months vs 3 years?), a cascading-failure mode that models how a shock in one asset propagates through related positions, and eventually a server-side deployment so the tool can run as a standalone product beyond the hackathon demo.

Built With

• anthropic
• api
• claude
• javascript
• json
• react
• search
• sonnet
• web