IdeaForge

Inspiration

We've both had ideas that died in a notes app. Not because they were bad, but because "I want to build X" never turned into "here's what I do today." Most idea tools either spit out generic task lists or skip straight to code without ever questioning if the idea makes sense. We wanted something that argues with you first — like a skeptical investor — before it ever hands you a plan.

What it does

IdeaForge takes a vague startup idea and runs it through an AI that acts like a cynical VC, asking 2-3 sharp questions that expose the riskiest unstated assumptions. Once you answer, it generates a full execution plan: an immediate first action you can take today, a 30/60/90 day roadmap, a risk matrix with mitigations, and an uncertainty rating so you know exactly how validated your idea actually is.

How we built it

We split the system into two layers. The backend runs a LangGraph state machine on FastAPI — Node 1 interrogates the idea, pauses for human input, and Node 2 takes the answers and generates a strictly typed roadmap using Pydantic-enforced JSON output via Groq/Llama 3. The frontend is React + Vite + Tailwind, built as a linear stage-based flow rather than a chat interface, so the experience feels like progress, not conversation.

Challenges we ran into

Getting Llama 3 to reliably return clean structured JSON was harder than expected — early outputs would wrap in markdown or drift from the schema, so we added Pydantic validation with a fallback parse step. We also had to resist scope creep constantly; we originally considered adding revenue estimates and "breakthrough point" predictions before realizing those would be the first thing judges challenge in Q&A, and cut them in favor of process-based outputs that are directionally honest instead.

Accomplishments that we're proud of

The clarification step actually feels adversarial in a useful way — testing it with ideas across very different domains, from technical products to more unconventional ones, it consistently asked questions we hadn't thought of ourselves. Getting a full LangGraph human-in-the-loop pipeline working end to end in under a week, with strict schema validation on the AI output, is something we're genuinely proud of given the timeline.

What we learned

We learned that the hardest part of "AI planning tools" isn't the AI, it's the handoff points — passing state cleanly between clarification and generation, and making sure the UI never shows the user a half-formed or malformed result. We also learned that constraining scope aggressively produces a better demo than trying to cram in every impressive-sounding feature.

What's next for IdeaForge

Right now IdeaForge ends at the plan. The natural next step is closing the loop between planning and execution — breaking the roadmap into atomic development tasks that get handed directly to AI coding agents, with a human reviewing each step before the next one fires. Effectively turning IdeaForge from a planning tool into the front door of a full idea-to-shipped-product pipeline.

Built With

• groq
• llm
• pydantic
• python
• react
• tailwind