Thesis

VC analysts drown in pitch decks. Thesis reads any PDF and returns a deal analysis, thesis fit score, investment memo, and pipeline tracking in under 30 seconds. Built for the other side of the table.

Comment

Inspiration

I'm 16 and an incoming venture capital analyst at 1435 Capital Management, a Princeton-based early-stage VC firm. I've built pitch decks for competitions and projects, and I spend a lot of time on LinkedIn talking to people in the startup and VC space. Before I even touch my first deal this summer, I already knew what the most painful part of the job would be.

Pitch decks.

So before writing a single line of code, I reached out to people in the VC space and asked directly: would a tool that reads pitch decks and returns structured analysis actually be useful? Every single person said yes. One told me they spend more time reading decks than making investment decisions. That was enough. I built Thesis.

When I saw this hackathon, I knew immediately I had to submit. It's built around exactly the problem Thesis solves. Coming from a non-tech background, building something I'm genuinely passionate about has been one of the best experiences I've had, and it's made me want to keep going and take this further.

The average VC firm sees 50+ decks a week. Each one takes around 30 to 45 minutes to read, extract insights from, score, and write up. 50 decks/week×40 min/deck=2,000 min/week 2,000÷60=33.3 hours/week

That’s 33+ hours of repetitive manual work every week before a single investment decision even gets made.

33.3×$75/hr≈$2,500/week $2,500×52≈$130K–$145K/year Per firm, that’s $130K–$145K annually spent just on deck processing labor.

Every tool built for pitch decks helps founders improve them. But nothing exists for the analyst on the other side of the table who has to process them at scale.

That gap is where Thesis lives.

What it does

Upload any pitch deck PDF. And in under 30 seconds, Thesis returns:

  • Structured deal data - market sizing, traction, team, business model, and asking terms pulled directly from the document
  • Thesis fit scoring - every deal is scored against your firm's specific criteria, with a quoted reason from the deck for each score
  • Red flag detection - no revenue model, uncited market claims, solo founder, unrealistic projections - surfaced automatically
  • AI content detection - section-by-section breakdown of how much of the deck appears AI-generated, with flagged excerpts and just plain English reasoning so it isn't confusing. Because analysts deserve to know what they are actually reading.
  • Investment memo - one page, structured for partner review, written to sound like a human analyst wrote it. Ready to share without editing.
  • Deal pipeline - Kanban board and sortable table tracking every deal from first look to final decision
  • Thesis configurator - each firm defines their own criteria with custom weights so every analysis reflects what that firm actually cares about, not just a generic rubric
  • Analyst notes - private per-deal notes that save automatically, so your thinking stays attached to each deal

How we built it

I used Next.js 13, TypeScript, Tailwind CSS, Framer Motion, Anthropic Claude API, Vercel. Built completely solo

The main insight was that Claude reads PDFs natively as document content blocks. No parsing libraries. No chunking. No preprocessing. The full deck goes in, structured JSON comes out.

The thesis configurator stores each firm's scoring criteria so every analysis is automatically calibrated to that firm's actual investment focus. The memo generation uses a separate prompt calibrated to produce something that sounds more like a senior associate wrote it, not just a chatbot summarizing a summary.

Challenges we ran into

Working solo was genuinely hard. Every bug, every broken page, every crash at 1am was mine to fix alone. There was no teammate to ask. Just me, the terminal, and whatever I could find online. There were also moments where the whole app broke from a single TypeScript error buried three files deep, and I had to trace it back from scratch.

The pitch deck problem was harder than I expected. Some decks are 8 slides. Some are 47. Some cite market size with a third-party source. Some just say "the market is big." I built fallback handling throughout, so if something cannot be extracted with confidence, it shows as "Not stated" rather than a made-up number. In VC, a confidently wrong number is way worse than no number at all.

Another challenge was getting the investment memo to read like a human wrote it took more iterations than anything else in this project. There is a real difference between an AI summary and something a partner would actually read on like a Monday morning. That difference really drove every single prompt revision until it cleared the bar that I wanted.

Accomplishments that we're proud of

Finishing this solo and building something I will genuinely use this summer, along with the other Venture Analysts at the firm.

The AI detection feature came directly from one conversation with someone who mentioned they had started noticing too many decks reading exactly the same way. And that one comment became an entire feature.

The red flag detection really surprised me. When you give the model a full deck and ask it to surface what a skeptical investor would flag, it consistently catches things a tired analyst on a Friday afternoon would for sure miss.

And the pipeline view actually feels useful rather than decorative. Sorting deals by thesis score and seeing red flag counts at a glance changes how you prioritize your week.

What we learned

Building for professionals is a completely different problem from building for regular users. Every single decision had to pass one test: would a VC partner open this on a Monday morning and immediately trust it? That bar shaped everything from the typography to the memo structure to how red flags get displayed.

Talking to users before writing a single line of code was one of the most important steps in this process. It directly shaped what I focused on, and even led to features I never would’ve thought of on my own. The AI detection feature, for example, came straight out of one conversation with a user. That wasn’t something I expected to learn going into a hackathon.

What's next for Thesis

Thesis is still early, but the direction is pretty clear. The goal is simple: take something that’s currently slow, manual, and repetitive in VC and make it fast, structured, and actually usable.

  • Live AI analysis at scale once API access is expanded, so decks get broken down in seconds instead of hours
  • Slack integration to push deal summaries straight into partner channels without anyone having to rewrite them
  • Comparable deal benchmarking to show how a startup stacks up against similar-stage companies
  • Notion and Linear integration to turn deal feedback into actual tasks and follow-ups automatically
  • Multi-analyst support for firms, with shared pipelines, notes, and role-based access
  • CSV export for LP reporting and internal tracking

This project is part of the Internal Tools Hacks hackathon on Devpost, which focuses on building tools that save time, cut costs, and improve real workflows inside companies.

That’s basically been the goal since day one.

If it does well in the hackathon, I’d absolutely love to use the prize money to build out these next features and move it closer to something real firms can actually use, not just a prototype.

Thesis started as a hackathon project, but the goal has always been to turn it into something that fits into real VC workflows, with the prize funding helping make that possible.

So yeah, the next step is simple: finish strong at the hackathon, and keep building until it actually matches how firms work day to day.

Built With

  • anthropic-claude-api
  • framer-motion
  • next.js
  • supabase
  • tailwind-css
  • typescript
  • vercel
Share this project:

Updates