Deepfield

Add a market screen
Settings screen
connect market screen
Opportunity simulator screen
Explore the market screen

Inspiration

I saw that founders rarely struggle because they have no ideas. They struggle because they don't know where to focus. As AI can help build products fast, a founder needs to know what it needs to build in first place.

DeepField was inspired by the idea of a map for markets - a way to explore industries/markets through their users, companies, pain points, technologies, signals, and underserved opportunities. Instead of generating random startup ideas, DeepField helps founders understand a market deeply and decide what is worth building.

## What it does

DeepField is an AI operating system for market research, opportunity discovery, and product strategy.

Users can:

Explore infinite markets such as Travel, Healthcare, or Artificial Intelligence etc
Navigate an interactive market graph
Expand individual nodes and explore deeper topics
View companies, users, pain points, solutions, signals, and opportunities
Add, remove, and switch between markets
Connect two markets to discover intersection opportunities
Compare Bootstrapped, VC, and Solo Builder paths
Review risks, assumptions, pricing, go-to-market, MVP steps, and a structured build stack
Export an opportunity plan as Markdown
Choose different AI providers and model-quality levels

Every opportunity receives a consistent score based on demand, pain severity, willingness to pay, whitespace, and feasibility:

$$ S = 0.30D + 0.25P + 0.20W + 0.15G + 0.10F $$

## How I built it

DeepField uses a modular frontend built with native JavaScript ES modules. The code is organized into feature, service, UI, core, and data layers.

The Node.js backend:

Orchestrates multiple AI providers
Validates structured JSON responses
Normalizes opportunity scores
Handles provider fallbacks and incomplete responses
Supports compact analysis and optional live research
Uses caching and request cancellation
Keeps API keys server-side for the hosted deployment

The Opportunity Simulator converts a market opportunity into a decision memo containing evidence, customers, pricing, risks, assumptions, MVP steps, founder paths, and a build stack.

GPT-5.6 was the primary reasoning and generation model throughout development. Codex acted as the repository-aware implementation partner by inspecting files, applying changes, refactoring code, running checks, debugging errors, configuring Vercel, and pushing working increments to GitHub.

## Challenges I ran into

AI providers returned structured responses differently, requiring provider-specific parsing and validation.
Rich AI responses were initially slow and expensive, so I introduced compact prompts, caching, model-quality controls, and cancellation.
Opportunity scores became inconsistent between the graph, opportunity feed, and simulator, so scoring was centralized on the server.
The simulator sometimes displayed stale data from a previous opportunity, requiring better request and active-market handling.
Vercel initially executed browser code as server code, causing a "document is not defined" crash. I fixed this by separating static assets from serverless API routes.

## Accomplishments that I am proud of

Built an end-to-end market exploration workflow instead of a simple chatbot.
Created an interactive graph that lets users explore markets visually.
Built a simulator that helps users decide how to pursue an opportunity.
Added three practical founder paths: Bootstrapped, VC Scale, and Solo Builder.
Added a structured build stack connecting discovery to execution.
Implemented centralized opportunity scoring.
Added multiple AI provider options and server-side API key handling.
Deployed a working version on Vercel.
Iterated on the product based on real testing feedback.

## What I learned

The most valuable AI output is not always the generated answer. It is the decision structure around the answer.

Founders need to understand:

What the opportunity is
Who experiences the problem
Who will pay
What assumptions are being made
How the score was calculated
What the risks are
What can realistically be built next

I also learned that perceived speed depends heavily on communication. Instant previews, visible progress, cancellation, compact responses, and clear fallback states make AI applications feel much more usable.

## What's next for DeepField

Features I am thinking of adding in future: