FOUNDERAI

Inspiration

Too many student startup ideas die quiet deaths after months of wasted engineering effort because no one asks the hard questions early enough. Traditional LLMs act as yes-men, consistently validating bad or vague assumptions out of conversational politeness. We wanted to build an institutional-grade, anti-complacency engine that acts as a strict, rubric-driven cross-examiner to help student founders either find a viable path or fail fast before wasting an academic semester.

What it does

FounderAI runs student founders through a structured, micro-segmented 10-stage track engineered directly around an ironclad, multi-tiered feasibility and impact framework.

Instead of holding a polite conversation, the system actively cross-examines assumptions to extract precise data metrics across 10 specific evaluation dimensions. If a user forces compilation prematurely without providing hard data, the engine applies an automatic early calculation penalty, capping all rubric scores at a maximum of 2 out of 5 points per node and tethering the total cumulative score between 20% and 35% maximum.

Upon clicking "Generate Roadmap," the engine performs an automated calculation and outputs a raw JSON telemetry payload containing a comprehensive assessment summary, precise structural evaluation breakdown text, and specific metric scores. This data populates a high-fidelity workspace modal that displays the calculated scores and equips the founder with three critical, binary operational directives:

1. Commit Node to Dashboard: Saves the project metadata to browser storage, updates the track status to "Validated," and locks down a live 10-step progress-tracked validation roadmap.
2. Dismiss: Closes the modal overlay and drops the user back into the active loop workspace to continue parameter refinement and clear conversational roadblocks.
3. Abandon Idea: A built-in anti-complacency exit trigger that updates the project status flag to "Abandoned" in storage, clears out current milestone indexes, and purges the unviable venture vector to protect the founder's time. Upon commitment to dashboard, the engine also generates a PDF summary of the idea and its evaluation based on the provision and evaluation of the conversation had on the main page. The PDF is sent to the Vault page where the user can then download it to their local storage.

How we built it

We engineered the platform using Next.js 14 and TypeScript to enforce a rigid, predictable type system. The front-end interface utilizes a geometric minimalist design aesthetic built with custom Tailwind CSS.

The reasoning layer leverages the hyper-fast Groq API running the llama-3.3-70b-versatile model. To bypass the UI layout breaks common with standard text streaming, we built a background data handler that intercepts the raw model completion. This handler uses a Dual-Layer Regex Meta-Scraper to target specific syntax boundaries (css-executable-thk ... and loose trailing JSON patterns). It strips the raw JSON telemetry metrics from the stream before rendering, extracting the status string and confidence integer to update interactive UI progress bars smoothly without causing screen jitter or layout shifts. All active project matrices and chat history strings are preserved locally via the Web Storage API (LocalStorage).

Challenges we ran into

Integrating real-time algorithmic scoring safely alongside free-form text chat proved highly difficult. Standard text streams often break structural UI layouts. To solve this, we built a robust background scraping protocol using regex boundaries that cleanly strips telemetry objects from the user-facing copy before it renders.

Balancing the fine line between helpful validation and encouraging rapid idea abandonment also required fine-tuning the prompt rulesets to ensure the AI grades stringently without discouraging genuine innovation.

Accomplishments that we're proud of

We successfully translated a complex, manual institutional matrix into an automated, responsive software system. We are proud of engineering the early compilation penalty mechanism, which proves the system can protect its own evaluation integrity when fed lazy or abstract data.

What we learned

We learned that AI can be configured to be an objective auditor rather than just a generative chatbot. Structuring prompts around hard mathematical rubrics completely shifts the value loop from open-ended text to actionable, binary decisions.

What's next for FounderAI

We want to integrate collaborative multi-founder profiles so student project teams can tackle roadmaps collectively. We also intend to expand our data mapping layers to automatically hook into live APIs for automated top-down market sizing verification.

Built With

• fetch
• google-gemini-1.5-flash-api
• groq
• groq-api-(llama-3.3-70b)
• jspdf
• localstorage
• lucide-react
• next.js-(react)
• next.js-14
• react
• tailwind-css
• typescript