ROAST
Startup Survival Simulation
MakersLounge #11 · Toronto Tech Week · Track 3: Synthetic Customers
Nazanin Ghelichi · Solo Build
Startup Survival Simulation
MakersLounge #11 · Toronto Tech Week · Track 3: Synthetic Customers
Nazanin Ghelichi · Solo Build
of startups fail — most because they built something nobody actually wanted.
average burned before a founder gets honest feedback from real users.
minimum for a real focus group. Slow, expensive, and still only 8 people.
Real customer studies are slow, expensive, and often fail to surface what customers actually want.
The track asked us to fix that.
ROAST does it in 60 seconds.
Min 140 characters — enough detail for the AI to understand the product, market, and value prop.
An LLM infers your target market and generates 4 real distinct people — names, ages, personalities, backstories. Nothing hardcoded.
Week 1 excitement, Month 1 habits, Month 3 truth. Each returns structured feedback: story, score, best/worst moment, willingness to pay.
Alternating roasts, damage scores, life bars. Real disagreement, not a consensus machine.
RAG-powered investor scores across 5 business dimensions + comparable startups + survival verdict. Then it roasts you.
Every agent runs on a different model with its own API key and independent memory context. No shared reasoning. No cross-contamination. The disagreement is genuine.
All 4 persona simulations run in parallel via concurrent API calls. Structured JSON output — stories, scores, moments — parsed with a fault-tolerant extractor that handles malformed responses.
The judge retrieves a startup evaluation rubric at inference time and scores across problem clarity, market size, feasibility, differentiation, and revenue potential.
Every LLM call has retry logic with fallback models. Casting tries 3 different models before failing. The app never crashes on a bad response.
Flask backend · Groq API · Chart.js · html2canvas PDF export · UUID result storage · Railway deploy
Instead of a text box, founders upload a full product brief — pricing, competitors, go-to-market, risks. LLMs react to real numbers, not guesses.
Synthetic users post async over 30 days — Week 1, Week 2, Month 1 updates. Opinions evolve. A living focus group, not a one-shot run.
LLM personas post to a real public thread. Real humans can stumble in and reply. Hybrid synthetic + human feedback loop.
Replace rubric.md with a vector DB of YC post-mortems, Crunchbase outcomes, and App Store retention benchmarks. Grounded in real startup data.
Built solo · One week · MakersLounge #11