AI/ML · Santa Clara University — GenAI Course
LitLens — Literature Intelligence Platform
Multi-agent research platform that synthesizes 10–50 papers into contradiction analysis, evidence scoring, and citation-ready literature reviews
The Challenge
PhD students and academic researchers spend weeks manually synthesizing 10–50 research papers — reading each one, tracking contradictions across studies, scoring evidence quality, and identifying literature gaps. No existing tool automates the full synthesis pipeline from PDF upload to citation-ready literature review draft.
My Role & Approach
Designed a multi-agent architecture where each agent specializes in one analytical task — ingestion, claim extraction, contradiction detection, methodology comparison, evidence scoring, gap analysis, and literature review generation. Made the key decision to use gpt-4o-mini across all agents for cost efficiency (~$0.01 per analysis) while parallelizing agent execution for speed.
Product Thinking
Problem Framing
Initial assumption was researchers needed 'better search across papers.' User interviews with PhD students revealed the real pain: not finding information, but synthesizing contradictions across 30+ papers and identifying what the field has missed. This shifted the product from search to automated synthesis.
Key Tradeoffs
8 specialized agents vs. 1 general agent: a single agent was simpler but produced inconsistent analysis quality across tasks. Chose multi-agent despite higher complexity because contradiction detection requires fundamentally different reasoning than gap analysis. Also chose gpt-4o-mini over gpt-4o — 10x cheaper, acceptable quality for structured extraction tasks.
What We Didn't Build
Scoped out citation graph visualization, collaborative annotation, and journal submission formatting. Each was requested by early users, but none solved the core problem: 'Help me understand what 30 papers say, where they disagree, and what's missing.' Staying focused on synthesis kept the product coherent.
Solution & Execution
Full-stack research platform with 8 LangGraph agents orchestrated in a parallel pipeline. Users upload PDFs via drag-and-drop, define their research question, and receive analysis across 6 tabs: Overview, Contradictions (with severity ratings), Methodology comparison, Evidence scoring (0–100), Gap analysis, and a thematic Literature Review draft. FAISS vector store with OpenAI embeddings powers a RAG chat for follow-up questions grounded in uploaded papers. React frontend with dark theme; FastAPI backend with 12-thread parallel ingestion.
Tech: LangGraph, LangChain, GPT-4o-mini, FAISS, React, FastAPI, OpenAI Embeddings, Python
Impact
- 8 specialized LangGraph agents running in parallel pipeline for comprehensive analysis
- ~$0.01–$0.07 per analysis run using gpt-4o-mini — accessible for student researchers
- 6 analysis tabs: contradictions, methodology, evidence scoring, gaps, literature review, and RAG chat
- Parallel paper ingestion across 12 threads for fast processing of large paper sets
Challenges & Pivots
Challenge: Single-agent approach couldn't handle the diverse analytical tasks — contradiction detection requires different reasoning than gap analysis
Resolution: Split into 8 specialized agents with LangGraph orchestration, running contradiction + methodology + evidence agents in parallel to reduce total analysis time.
Challenge: Follow-up questions about papers required re-reading entire documents each time
Resolution: Built FAISS vector store with text-embedding-3-small to index all uploaded papers, enabling grounded RAG chat that retrieves relevant passages for each question.
Screenshots
What I Learned
- Multi-agent architectures shine when each subtask requires fundamentally different reasoning patterns — specialization beats generalization.
- Cost-per-query matters for academic tools — gpt-4o-mini at $0.01/analysis made the tool accessible to students who can't afford expensive API calls.
- Parallel agent execution is critical for user experience — sequential 8-agent pipelines would be unusably slow.
Next case study
Lending Club Loan Default Prediction