sanyam.
PDFarXivpreparestrip bibread_paperfits? one callmap sectionsentities, resultssynthesizefindingsgradeLLM judgeweak → retry onceanalysis13 entities18–34 resultsprovider cascade · any error or rate limit moves onGemini FlashprimaryGroq Llama 3.3overflowOllama Qwen2.5offline~3 LLM calls per paper, down from ~16pgvector graph: extends, replicates, contradicts, shares-method
PDFarXivmap sectionsin parallelsynthesizefindingsgradeLLM judge, retry once3 calls per paperdown from 16Gemini, then Groq, then Ollama. ROUGE-1 F1 0.64.
All projects

PaperMind

Faithful Research-Paper Intelligence

Abstract-only summarizers miss what a paper actually shows. PaperMind processes the full PDF, every section, table, and result, then grades its own output for faithfulness before returning it.

Result
ROUGE-1 F1: 0.64 vs abstracts
90/90 tests passing
$0 API cost (free tiers)
What it does

LangGraph-orchestrated pipeline that turns full academic PDFs into structured intelligence: map-reduce summaries, typed entity and results extraction, LLM-judge quality grading, and a knowledge graph of how papers relate.

What I built
  • LangGraph map-reduce pipeline covers the whole paper: every section is summarized concurrently, then synthesized into a 300 to 450 word summary with findings, contributions, and limitations.
  • An LLM-as-judge grading node scores faithfulness and specificity and retries weak results automatically, with cascading failover across Gemini 2.0 Flash, Groq Llama 3.3, and local Ollama Qwen2.5.
How it works
  1. 01React 18 + Vite Frontend
  2. 02FastAPI (async)
  3. 03LangGraph Engine
  4. 04Supabase (pgvector)
  5. 05Gemini 2.0 Flash
  6. 06Groq Llama 3.3
  7. 07Ollama Qwen2.5 (cascading failover)
Features
  • Map-reduce summarization: prepare, map sections concurrently, synthesize, then grade
  • Typed entity and quantitative-results extraction via Pydantic schemas, from prose and real PDF tables; no regex
  • LLM-as-judge quality control: weak summaries loop back and get retried automatically
  • RelationAgent labels paper-to-paper links (extends, replicates, contradicts, shares-method) into a pgvector knowledge graph
  • Domain-agnostic across biomedicine, physics, and ML papers
  • Benchmark harness runs the full pipeline against real arXiv papers: 100% success rate, up to 2.5x parallel speedup, and it caught a real figure-extraction bug
  • Runs entirely on free LLM tiers (Gemini, Groq, local Ollama) with rate limits absorbed by concurrency caps and backoff retries
Previous
BEDROCK
Next
FlowSync