PitchLab

Inspiration

Ideas are easy.

Anyone can come up with one at any time.

But early-stage startups don't fail because the ideas are bad; they fail because founders lack a clear way to scale them in the chaos of day one.

We built PitchLab to bridge that gap. We take your raw idea and stress-test it through a virtual board of experts, turning unpolished thoughts into structured, investor-ready feedback so you can stop guessing and start building.

Features

Multi-Agent: problem, market, monetization, and tech agent specialists run in parallel; specialization means sharper, more accurate output than any single prompt
Real-Time Market Data: agents query Tavily Search during runtime, so feedback reflects today's market, not pre-existing training data
Built-In VC Frameworks: agents reason with credible VC frameworks including Hamilton Helmer's 7 Powers, Paul Graham's frameworks, SaaS unit economics, and more
Live Dashboard: confidence scores update in real time as agents scrape market data and process your follow-up answers
Voice Pitch: record your pitch and get a live delivery score (pacing, fillers, pauses) via ElevenLabs (or just type)
Published on ASI:1 Agentverse: agents are live and discoverable, so any company can plug our validate into their own product

Tech Stack

Frontend: React + Vite, agent cards update live via Server-Sent Events as results stream in, not as a single batch

Backend: FastAPI with fully async request handling, structured JSON contracts per agent, and real-time Tavily Search injected mid-run into each agent's context window

AI Layer: 4 specialized agents run in parallel via asyncio.gather() — each scoped to a single domain (problem, market, monetization, tech) with embedded VC frameworks baked into the system prompt. DeepSeek drives inference. ElevenLabs Scribe v2 handles voice transcription for delivery scoring.

Agent Infrastructure: Agents are published on Fetch.ai's Agentverse via the uAgents framework independently discoverable by external teams

Challenges We Ran Into

One of the primary challenges was ensuring that the system produced outputs that were both specific and actionable. Early iterations generated responses that were coherent but overly generic. Addressing this required careful design of agent roles (i.e. what metrics do startup evaluators like YC Combinator care about?), prompt structures, and output formats.

Another challenge involved managing inter-agent inconsistency. Because each agent evaluates a different dimension, conflicting conclusions frequently emerge. Instead of suppressing these conflicts, we implemented a “Q&A” stage that embraces conflict; it asks clarifying questions to better understand vague areas, improving the overall reliability of the system.

With most computation happening asynchronously in the backend, we had to design a UI that felt alive and responsive, not like a loading screen. We debated heavily on how to surface agent progress in real time without overwhelming the user, undergoing multiple revisions.

Accomplishments That We're Proud Of

Key accomplishments include:

Supporting both natural voice (through ElevenLabs) and text pitches to enhance accessibility
Moves beyond a harsh "pass/fail" model to provide a roadmap of weaknesses and fixes
Functions as both an investor’s decision tool and a founder’s learning platform
Educated agents scrape the web and cite credible sources to eliminate hallucinations
Adaptively inquires whenever confidence scores are low to ensure the final feedback is grounded in the user’s vision (not just AI)

Most importantly, we are proud to have built a system that explores a recently emerging paradigm: multi-agent reasoning.

What We Learned

This project reinforced that the effectiveness of AI systems depends not only on the underlying models, but on how multiple systems can be orchestrated.

We learned that:

Presentation matters just as much as code
Structured decomposition of problems leads to higher-quality outputs
Pointing out uncertainty improves user trust and decision-making
Multi-agent systems can better simulate real-world reasoning than single-model approaches and prevent hallucination
Another thing: Generative AI has obviously changed the game for hackathons and the industry. It's important to not get lost in “prompting hell”. Never blindly press the tab. If you hit a bug in code you don’t understand, the project is doomed. We learned that the hard way…

What's Next for PitchLab

Our next steps focus on expanding both the depth and accessibility of the platform.

Planned improvements include:

Enhanced agent reasoning with richer real-world data integration
Persistent idea tracking to support iterative development
More advanced voice interaction using ElevenLabs
Personalization based on user goals and experience level
Collaboration features for teams and classrooms

In alignment with the Light the Way track, we envision PitchLab evolving into an educational platform that empowers students and early-stage builders to develop structured thinking skills. By making expert-level reasoning accessible, we aim to support more inclusive and effective learning environments.

Our long-term goal is to establish PitchLab as the standard entry point for idea evaluation, enabling anyone to transform ideas into informed, actionable decisions. :)