Uddyog-CoPilot

Inspiration

In Bangladesh and across emerging markets, thousands of early-stage founders fail not because of bad ideas, but due to execution fog. They lack structured guidance, objective feedback, and actionable validation frameworks. Most existing tools either oversimplify business advice into generic lists or overwhelm founders with walls of text.

We were inspired to build Uddyog-CoPilot—a localized, Bangla-first AI startup co-pilot that acts as a real-time execution partner, bridging the gap between uncertainty and meaningful action.

What it does

Uddyog-CoPilot transforms early-stage startup friction into a structured execution roadmap across four core pillars:

• War Room & Business Guide: Runs a multi-agent AI debate using distinct personas (Skeptic, Strategist, Operator) to stress-test business ideas and map risks into an interactive Assumption Map.
• Task Manager & To-Do List: Converts high-level startup strategy into actionable, day-to-day operational tasks for founders.
• Launchpad & Strategy Room: Drafts targeted customer outreach campaigns and leverages document-based AI advisors to answer complex strategic queries from uploaded pitch decks.
• Live Pitch Session: Delivers real-time, interactive pitch coaching using camera, mic, and multimodal streaming feeds.

How we built it

We built Uddyog-CoPilot with a modern full-stack tech architecture:

• Frontend: Next.js 15 (App Router), React 19, TypeScript, Tailwind CSS 4, and Framer Motion.
• AI Orchestration: Google Gemini Flash for multi-agent reasoning and Gemini Live Multimodal APIs for real-time video/audio pitch coaching.
• Backend: Express 5 REST API powering session persistence and vector retrieval pipelines.
• Database & Vector Search: PostgreSQL via Prisma 6 ORM paired with pgvector for indexing business documents.
• Monetization & Web3: Integrated x402 Micropayment Protocol middleware for handling pay-per-use premium AI analysis.

Challenges we ran into

• Enforcing Genuine AI Disagreement: Preventing AI agents from falling into false consensus required deep system prompt engineering so that the Skeptic, Strategist, and Operator maintain distinct, opposing frames of thought.
• Context-Aware Document Retrieval: Optimizing PDF chunking and vector embedding search via pgvector so the AI advisor returns instant, relevant insights from uploaded business plans.
• Low-Latency Multimodal Streaming: Synchronizing real-time camera overlays and audio streams via Gemini Live WebSockets with minimal latency during pitch sessions.

Accomplishments that we're proud of

• Agentic Reasoning Pipeline: Successfully engineered a multi-agent debate structure that surfaces hidden startup risks automatically.
• Bangla-First Native UX: Created an intuitive, localized user experience tailored specifically to empower regional entrepreneurs.
• Full-Stack & Web3 Integration: Successfully integrated vector search, Web3 micropayments (x402 protocol), and WebSocket streaming within a tight timeline.

What we learned

• How to design agentic workflows that focus on reasoning and tradeoff analysis rather than basic text generation.
• Implementing low-latency vector databases directly inside PostgreSQL using Prisma and pgvector.
• Balancing automated AI guidance with Human-in-the-Loop control points—ensuring the AI surfaces evidence while the founder retains final decision-making power.

What's next for Uddyog-CoPilot

• Live Local Market Intelligence: Connecting real-time regional web search to validate target market sizes automatically.
• Expanded Micropayments: Scaling the x402 protocol integration to allow seamless pay-per-use access to specialized market reports.
• Mobile Companion App: Bringing daily To-Do lists and quick voice pitch practice directly to iOS and Android mobile platforms.

Built With

• ai
• copilot
• javascript
• python
• replit