Lynx | Devpost

Inspiration

Canada has world-class engineering talent. The corridor between Waterloo and Toronto produces some of the best technical researchers and student builders in the world. But we have a massive brain drain problem. We aren't losing founders to the US because Canada lacks capital. The money is here. We lose talent because the friction between student builders, IP holders, and local capital is too high.

Academic researchers and student engineers speak in technical documentation, GitHub commits, and theoretical breakthroughs. Institutional investors speak in term sheets, burn rates, and market fit. Because this translation layer is broken, US firms often swoop in, acquire early-stage Canadian IP, and pull the talent south. We wanted to fix this friction. We built a platform that translates technical IP into an institutional thesis and connects Canadian founders directly with the right local funds.

What it does

Lynx is a platform that helps technical founders find and pitch to the right investors.

IP Processing: It reads technical documents or academic papers and extracts the core commercial utility.
Sourcing: It scrapes live directories of venture funds and structures the HTML into a clean database using the Gemini API.
Matchmaking: It matches the startup's technology against the specific investment mandates of Canadian funds.
Pitch Simulation: It creates an interactive audio simulation of the matched VC, allowing the founder to practice their pitch and answer realistic pushback before the actual meeting.

Build Process

We focused on API orchestration and strict logic pipelines.

Frontend: We used Google Antigravity to quickly build a React interface where founders can upload documents and review their pitch materials.
Backend: We built a FastAPI Python server.
Adversarial Matchmaking & Data Analysis: We used Backboard.io to handle document storage and memory using their deep document RAG & stateless memory system. To determine if a startup is a good fit for a VC, we set up a multi-agent system. One agent acts as the startup advocate, and another acts as the VC critic. They evaluate the startup's tech against the VC's mandate. A third agent scores the interaction using a formula that weighs vector similarity against stage mismatch and debate performance:

M(S, V) = α * sim(v_S, v_V) - β * Δ_stage - γ * P_debate

The Simulation: We used the Gemini API to extract data and handle the simulated VC logic. We connected this text output to the ElevenLabs API to stream real-time audio back to the user during the pitch practice.

Challenges we ran into

Scraping issues: Standard CSS selectors failed on inconsistent VC websites. We decided to pass raw HTML to the Gemini API to format it into JSON.
Logic loops: Getting two AI agents to debate in Backboard.io sometimes caused infinite loops where they would argue endlessly. We had to refine our system instructions to enforce strict turn limits and force a judge agent to output the final score.

What we learned

We learned that the real value of large language models is not just generating text, but orchestrating unstructured data and managing state across different tasks.