ADIA — Autonomous Decision Intelligence Agent

Investor-grade pitch analysis for every founder — powered by Amazon Nova multimodal AI on AWS Bedrock

Comment

Inspiration

I'm a student in India. I've watched brilliant founders around me pitch investors and get rejected — with zero feedback. No explanation. No second chance.

Professional due diligence costs thousands of dollars. Accelerator mentors are overbooked. Most AI tools just summarize what you already told them.

I built ADIA to fix that.

What It Does

ADIA is a multimodal agentic decision intelligence system powered by Amazon Nova on AWS Bedrock.

Upload a pitch deck PDF. ADIA converts each page to a PNG image and sends the actual visual content — charts, tables, traction graphs — to Amazon Nova Lite for analysis. Not extracted text. Real multimodal understanding.

A 4-agent pipeline runs in parallel and returns a structured verdict in under 10 seconds:

  • Verdict: GO / NO-GO / CONDITIONAL
  • Conviction Score: 0–100
  • Fatal Flaw: Flagged by Nova tool use — severity: critical / major / minor
  • Evidence Citations: Every point tagged [Page N] or [AI inference]
  • Similar Cases: Top 2 historical pitches from FAISS semantic search
  • Next Action: One concrete step the founder should take

How I Built It

Amazon Nova Integration — 3 distinct capabilities:

1. Nova Lite Multimodal Reasoning PDF pages converted to PNG images and sent directly to Nova Lite in the content array alongside text. Nova sees what an investor sees — charts, tables, visual data.

2. Nova Tool Use — Real Agentic Behavior The ReasoningAgent calls Nova with a flag_fatal_flaw tool defined in toolConfig. Nova decides when to invoke it — returning structured JSON with flaw, severity, and evidence. This is genuine tool use, not a prompt asking for a list.

3. Nova Multimodal Embeddings + FAISS After verdict generation, ADIA embeds the pitch and searches a local FAISS index of historical pitches — Airbnb 2009, Uber seed, Stripe, WeWork, Theranos — returning the 2 most similar cases with match % and lesson.

Architecture:

  • ResearchAgent + AnalysisAgent run in parallel via ThreadPoolExecutor
  • Results feed into ReasoningAgent + ReportAgent
  • Verdict streams token-by-token via converse_stream SSE
  • Total: 2 Nova calls, under 10 seconds end-to-end

Challenges

1. Multimodal PDF processing PyPDF2 only extracts text. Getting Nova to read actual page images required converting PDFs to PNG with pdf2image and sending raw image bytes in the content array. The visual analysis quality difference was immediate and dramatic.

2. Real tool use vs fake tool use Most "agentic" demos just ask the LLM to return JSON. Real tool use means defining tools in toolConfig and handling stopReason == "tool_use" responses. Getting Nova to reliably invoke flag_fatal_flaw with structured evidence took significant prompt engineering.

3. Latency 4 sequential Nova calls = 25+ seconds. Solution: merge agents into 2 calls, run in parallel with ThreadPoolExecutor, cap maxTokens at 300, add performanceConfig latency optimization. Result: under 10 seconds.

4. Render cold starts Free tier backend sleeps after inactivity. Solution: keep-warm /ping endpoint + UptimeRobot pinging every 5 minutes. Zero cold starts during demo.

What I Learned

Nova's multimodal API is simpler than expected — sending image bytes alongside text just works, and the quality difference vs text-only is dramatic.

Tool use changes how you think about agents. When Nova calls flag_fatal_flaw and returns structured JSON with evidence, it stops feeling like a text generator and starts feeling like a system making decisions.

Streaming matters more than I thought. Watching tokens appear changes how users engage with the output entirely.

Community Impact

India produces more engineering graduates than any country on earth. A fraction start companies. Almost none have access to quality investor feedback before their first pitch.

ADIA is designed as infrastructure for student incubators. Any university accelerator can deploy it and evaluate hundreds of student pitches with consistent, evidence-grounded analysis. Cost per analysis on Nova Lite: fractions of a cent.

The vision: a platform that university programs, startup weekends, and accelerators in emerging markets white-label and run — giving every student founder the preparation previously reserved for founders with the right network.

Built With

  • amazon-nova
  • aws-bedrock
  • faiss
  • fastapi
  • next.js
  • pdf2image
  • python
  • react
  • recharts
  • render
  • vercel
Share this project:

Updates