Paper2Proto

Home Page -> Attach Research Paper
Paper Summarization -> Summarize AI Agent
Research Paper Structure Visualization
Theoretical Research Implementation -> Builder AI Agent
Validation of Generated Content -> Verifier AI Agent

Inspiration

Reading ML papers often feels like detective work: nice equations, but missing details. We wanted a tool that behaves like a grounded, tireless research assistant instead of a generic, hallucination‑prone chatbot.

What it does

Paper2Proto turns PDFs into a conversational, evidence‑grounded workspace.
You upload a paper, and in chat you can:

Ask for summaries, clarifications, and derivations.
Get code‑oriented “protos” for experiments.
See answers that are verified against the original text, not invented.

How we built it

Frontend (React): Upload UI (“Evidence Locker”) and chat interface calling FastAPI endpoints.
Backend (FastAPI): Receives PDFs, stores them in Amazon S3, parses and chunks text, embeds via Amazon Bedrock, and indexes vectors + metadata into Amazon OpenSearch.
LangGraph agents (inside the backend):
- Router: Chooses task type (SUMMARY / QA_TUTOR / BUILD / MIXED).
- Retriever: Builds an Evidence Pack from OpenSearch.
- Producers: Summarizer, Tutor/Q&A, and Builder agents powered by Bedrock (and future providers).
- Verifier: Checks drafts against the Evidence Pack and drives a verify → re‑retrieve → regenerate loop.

Challenges we ran into

Designing a verification loop that improves answers without adding too much latency.
Keeping all agents strictly grounded in the paper instead of drifting into plausible fiction.
Handling messy, real‑world PDFs for chunking and retrieval.
Structuring the LangGraph so that agents are specialized but not overly complex.
Abstracting the LLM layer so we can mainly use Bedrock now but plug in others later.

Accomplishments that we're proud of

Moving beyond “chat with your PDF” to a clear, inspectable agent graph: Router → Retriever → Producer → Verifier.
Implementing an Evidence Pack + Verifier loop that actually corrects itself instead of just apologizing.
A Builder agent that proposes code and configs while citing the exact passages it relied on.
A modular architecture (FastAPI + S3 + OpenSearch + Bedrock + DynamoDB + LangGraph) that we can evolve without rewrites.

What we learned

Grounding is a pipeline property: indexing, retrieval, agents, and verification all matter more than a single clever prompt.
Retrieval quality dominates everything downstream; better chunks and metadata beat over‑tuned prompts.
Explicit graphs (via LangGraph) make debugging and extending agentic systems far easier.
Users trust honest “not found in the paper” answers more than confident hallucinations.

What's next for Paper2Proto

Multi‑paper Evidence Packs for literature reviews and comparisons.
Stronger Builder flows that generate end‑to‑end experiment scaffolds while staying faithful to the paper.
Better handling of math and figures, with links from questions to the exact equations/sections.
Human‑in‑the‑loop tools to inspect evidence, adjust reasoning, and feed back preferences.
A provider‑agnostic LLM layer so Bedrock, Gemini, and future models can power different agents without changing the graph.