FounderLens

Inspiration

As a student on an entrepreneurship exchange program myself, I realised that the most effort and time is spent on brainstorming, researching and validating ideas, but despite this, it is also when many crucial factors can be overlooked. Most of us also spend time researching on the internet and then validating with the people around us, but what if we could design an AI assistant that keeps track of, researches and critiques your idea all in one?

What it does

FounderLens gives founders access to three AI advisors, each with a distinct persona and evaluation framework:

The Investor assesses fundability, traction, deal-breakers. You can use it as a practice tool for actual pitches.
The Researcher grounds your assumptions in data, market size, competitive landscape, demand signals, and what's still unverified
The Co-Founder challenges your framing, untested assumptions, pivot opportunities, and the questions you haven't asked yet

Each agent holds its own conversation and persists between sessions. You can also use Multi-Agent Mode to submit your idea once and get all three perspectives in parallel, side by side. Every session produces a feasibility score (1–10). The Insights Dashboard tracks how that score evolves across sessions for the same idea, so you can see whether your refinements are actually moving the needle.

How we built it

The frontend is built in Next.js 16 with React 19, using Zustand for global state and inline styles throughout to sidestep Tailwind compilation issues on Apple Silicon. Each agent's theme is driven by a single AgentConfig object, keeping all design decisions in one place.

The AI layer uses Google Gemini 2.5 Flash via @google/genai. Every chat message triggers two concurrent Gemini calls:

A conversation response — the agent replies in character using its system prompt
A structured analysis — a second call extracts a strict JSON object containing feasibility_score, startup_killers, investor_concerns, recommendations, and suggestions for the next prompt chips

In Multi-Agent Mode, all selected agents run their calls concurrently viaPromise.all, so results stream in as each agent finishes. Conversations are persisted per agent in localStorage. Score evolution is computed by normalising idea titles into a ideaKey and grouping sessions sparkline.

Challenges we ran into

Prompt engineering depth. Getting each agent to stay genuinely in character, asking one question at a time, refusing to fabricate data, flagging assumptions rather than filling them in, required multiple rounds of iteration. The investor in particular had a tendency to be encouraging by default, which defeated the purpose.

Accomplishments that we're proud of

Multi-Agent Mode with live streaming results. Submitting one idea and watching three independent AI advisors respond in parallel , getting the concurrent state updates right so results render independently.

What we learned

How to write system prompts that produce consistent, structured personas
How to extract structured JSON from an LLM reliably alongside a freeform conversation response
How to design for agents with meaningfully different viewpoints