How I Built a Multimodal AI Investment Analyst for Student Founders Using Amazon Nova
A solo student built ADIA — an agentic decision intelligence system that reads pitch deck images, runs parallel AI agents, and delivers investor-grade verdicts using Amazon Nova Lite, multimodal embeddings, and tool use on AWS Bedrock.
The Problem No One Talks About
Every year, thousands of student founders pitch investors and get rejected.
Most never learn why.
Most never learn why.
Not because their ideas are bad. Because they had no way to stress-test
their pitch before walking into that room. Professional due diligence costs
thousands of dollars. Accelerator mentors are overbooked. And most AI
chatbots just summarize what you already told them.
their pitch before walking into that room. Professional due diligence costs
thousands of dollars. Accelerator mentors are overbooked. And most AI
chatbots just summarize what you already told them.
I built ADIA — Autonomous Decision Intelligence Agent — to change that.
ADIA is a multimodal agentic system powered by Amazon Nova on AWS Bedrock.
It reads your pitch deck as actual images, runs a 4-agent reasoning pipeline,
flags fatal flaws using real tool use, and delivers a structured verdict:
GO, NO-GO, or CONDITIONAL — with evidence, citations, and next actions.
It reads your pitch deck as actual images, runs a 4-agent reasoning pipeline,
flags fatal flaws using real tool use, and delivers a structured verdict:
GO, NO-GO, or CONDITIONAL — with evidence, citations, and next actions.
Built by one student. Free to use. Live right now.
Why Amazon Nova — Not Just Any LLM
Most pitch analyzers extract text from PDFs and feed it to a language model.
That misses everything visual — charts, traction graphs, financial tables,
team slides.
That misses everything visual — charts, traction graphs, financial tables,
team slides.
Amazon Nova Lite natively accepts images. So ADIA converts each PDF page
into a PNG and sends the actual visual content to Nova for analysis. When
a founder uploads their pitch deck, Nova sees what an investor sees — not
just the words, but the data visualized on the page.
into a PNG and sends the actual visual content to Nova for analysis. When
a founder uploads their pitch deck, Nova sees what an investor sees — not
just the words, but the data visualized on the page.
This is the core insight that makes ADIA different: multimodal
understanding is not a feature, it is the foundation.
understanding is not a feature, it is the foundation.
Beyond image understanding, ADIA uses three distinct Nova capabilities:
1. Nova Lite with Tool Use
The ReasoningAgent calls Nova with a
the toolConfig. Nova decides when to invoke it — passing a structured
JSON object with the flaw description, severity level (critical / major /
minor), and the evidence it found. This is genuine agentic behavior, not
just a prompt that asks for a list.
The ReasoningAgent calls Nova with a
flag_fatal_flaw tool defined in
the toolConfig. Nova decides when to invoke it — passing a structured
JSON object with the flaw description, severity level (critical / major /
minor), and the evidence it found. This is genuine agentic behavior, not
just a prompt that asks for a list.
2. Nova Multimodal Embeddings + FAISS
After generating a verdict, ADIA embeds the pitch using Nova's embedding
model and searches a local FAISS index of historical pitches — Airbnb's
2009 rejected deck, Uber's seed round, Theranos, WeWork, Stripe. It
surfaces the 2 most similar cases with match percentage and the lesson
learned. Founders see not just their verdict but where they fit in history.
After generating a verdict, ADIA embeds the pitch using Nova's embedding
model and searches a local FAISS index of historical pitches — Airbnb's
2009 rejected deck, Uber's seed round, Theranos, WeWork, Stripe. It
surfaces the 2 most similar cases with match percentage and the lesson
learned. Founders see not just their verdict but where they fit in history.
3. Parallel Agent Pipeline with Streaming
ADIA runs a 4-agent pipeline — Research, Analysis, Reasoning, Report —
but merges them into 2 parallel Nova calls using ThreadPoolExecutor.
The result streams back to the frontend via Server-Sent Events using
instead of staring at a loading spinner.
ADIA runs a 4-agent pipeline — Research, Analysis, Reasoning, Report —
but merges them into 2 parallel Nova calls using ThreadPoolExecutor.
The result streams back to the frontend via Server-Sent Events using
converse_stream. Founders watch the reasoning appear in real time
instead of staring at a loading spinner.
The Architecture
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
Pitch Text + PDF Pages (PNG images)
│
▼
┌──────────────────────┐
│ FastAPI Backend │
│ │
│ ┌────────────────┐ │
│ │ ResearchAgent │──┼──▶ Nova Lite Call #1
│ │ AnalysisAgent │ │ (parallel, multimodal)
│ └────────────────┘ │
│ │ │
│ ┌────────────────┐ │
│ │ ReasoningAgent │──┼──▶ Nova Lite Call #2
│ │ ReportAgent │ │ (tool use: flag_fatal_flaw)
│ └────────────────┘ │
└──────────────────────┘
│
▼
Nova Embeddings → FAISS → Similar Cases
│
▼
Streaming verdict → Next.js frontendTech stack:
- Backend: FastAPI, Python, boto3, aws-bedrock-runtime
- Frontend: Next.js, React, Recharts
- Vector search: FAISS (local, no external API)
- File processing: pdf2image, Pillow
- Model: amazon.nova-lite-v1:0 via AWS Bedrock
What the Output Looks Like
When a founder uploads their pitch, ADIA returns:
- Verdict: GO / NO-GO / CONDITIONAL
- Conviction Score: 0–100 with visual ring
- Fatal Flaw: flagged by Nova's tool use call, with severity badge
- Asymmetric Upside: what could make this a breakout
- Key Risks: each tagged [Page N] or [AI inference]
- Key Assets: evidence from the actual document images
- Similar Cases: 2 historical pitches from FAISS with match % and lesson
- Next Action: one concrete step the founder should take
A founder in Hyderabad with no advisor, no mentor, no network — gets the
same quality of structured analysis that a Series A investor runs
internally. In under 10 seconds.
same quality of structured analysis that a Series A investor runs
internally. In under 10 seconds.
The Community Impact
India produces more engineering graduates than any country on earth.
A fraction of them will start companies. Almost none of them will
have access to a quality investor before their first pitch.
A fraction of them will start companies. Almost none of them will
have access to a quality investor 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, structured, evidence-grounded analysis. The cost per analysis
on Amazon Nova Lite is fractions of a cent.
accelerator can deploy it and evaluate hundreds of student pitches with
consistent, structured, evidence-grounded analysis. The cost per analysis
on Amazon Nova Lite is fractions of a cent.
The vision is not a consumer app. It is a platform that university
programs, startup weekends, and accelerators in emerging markets can
white-label and run — giving every student founder the preparation
that was previously reserved for founders with the right network.
programs, startup weekends, and accelerators in emerging markets can
white-label and run — giving every student founder the preparation
that was previously reserved for founders with the right network.
What I Learned Building This
Three things surprised me during this build:
Nova's multimodal API is simpler than expected. Sending image bytes
in the content array alongside text just works. The model understands
charts and tables without any special prompting.
in the content array alongside text just works. The model understands
charts and tables without any special prompting.
Tool use changes how you think about agents. When Nova calls
feeling like a text generator and starts feeling like a system that
is making decisions. That shift in architecture changes what you can build.
flag_fatal_flaw and returns structured JSON with evidence, it stops
feeling like a text generator and starts feeling like a system that
is making decisions. That shift in architecture changes what you can build.
Streaming matters more than I thought. The difference between
waiting 8 seconds for a result and watching it appear token by token
is not cosmetic — it changes how the user interprets the output.
They read as it arrives. They engage differently.
waiting 8 seconds for a result and watching it appear token by token
is not cosmetic — it changes how the user interprets the output.
They read as it arrives. They engage differently.
Try It
Upload any pitch deck PDF or paste a pitch description.
Try the demo scenarios to see GO, NO-GO, and CONDITIONAL verdicts.
Try the demo scenarios to see GO, NO-GO, and CONDITIONAL verdicts.
Built with Amazon Nova on AWS Bedrock. Solo student submission
for the Amazon Nova AI Hackathon.
for the Amazon Nova AI Hackathon.
ADIA
Autonomous Decision Intelligence Agent
Powered by Amazon Nova
Autonomous Decision Intelligence Agent
Powered by Amazon Nova
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article