IdeaSense AI

IdeaSense began as a 10-week university industry capstone — a team project with fortnightly reviews with a mentor from Theta — after which I rebuilt it solo from the ground up. The current product shares little with the original team version beyond the problem space, and it is the clearest example of my AI workflow work: controlled orchestration, persistent state, review gates, and practical deployment trade-offs.

2025 — Present

An AI workflow system for structured early-stage venture evaluation, designed to turn open-ended founder interviews into confirmed, traceable evidence for DVF scoring and report generation.

Back to case studies Visit live product

Overview

IdeaSense AI guides early-stage founders through a 30–60-minute DVF assessment. The product does not treat chat as the final interface; it uses staged conversation, explicit confirmation, structured state, scoring, and report generation so each assessment can be reviewed and reused.

Problem

Founder interviews contain useful signal, but raw conversation is hard to compare or trust. Generic chat can drift, merge assumptions with facts, and produce confident output without a stable evaluation path. The challenge was to keep conversation natural while making the assessment process controlled and reviewable.

Role

AI workflow lead on the original team capstone, and sole builder of the current rebuild. I own the question-bank structure, orchestration logic, context design, model routing, verification strategy, and the path from prototype to a usable product flow.

Context and constraints

The original MVP had to fit a 10-week capstone schedule with prototype, MVP, and final-release checkpoints; the current system is a later solo rebuild.

The deployment target favored lightweight infrastructure, so the system could not depend on heavy local models or overly complex retrieval layers.

Latency and cost mattered, which forced trade-offs between fast conversational turns and stronger reasoning for summaries, scoring, and reports.

Non-technical founders needed the flow to feel simple, while the outputs still had to be explicit enough for review.

What I built

A DVF question bank and staged interview flow covering problem, market, technology, and report synthesis.

A deterministic turn orchestrator that replaced an overly flexible agent flow with explicit stage confirmation.

A layered context strategy using system rules, rolling summaries, recent turns, sanitized input, and stage-specific instructions.

Verification and evidence-handling hooks for claims, scoring, and final report generation.

Technical approach

Implemented a decoupled Next.js frontend and FastAPI backend with PostgreSQL state and SSE chat streaming.

Persisted structured workflow state instead of relying on raw chat alone, using confirmed fields, meta state, and pending-review values to separate facts from suggestions.

Used task-based model routing so conversational turns could stay fast, while stage summaries, scoring, and report generation could use stronger reasoning paths.

Extended the product with auth, project workspaces, sample report flows, export, and organization/admin screens.

Outcomes

Built a full end-to-end MVP with staged workflow control, persistent project state, SSE-based streaming, structured evidence capture, DVF scoring, and exportable reports.

A visible confirmation step (for example market→tech) initially took around 96 seconds; I attacked the perceived wait by moving non-blocking work — summaries, scoring, and report preparation — off the synchronous turn into background jobs, treating latency as a path-design problem rather than a model-speed one.

The main product lesson was that controlled workflow design made the AI output easier to trust than an unconstrained chat interface.

Reflection

Reliability improved when I stopped trying to make the agent more autonomous and made the workflow more explicit. The state-machine approach was less flashy, but it gave the product clearer boundaries and more defensible outputs.