Project Information
- Category: AI & Research
- Collaborators: 2 PhD researchers
- Timeline: 2024 – Present
- Tech Stack: LangChain, Python, Polygon, IPFS, Pinata, NestJS, RAG
Overview
A LangChain-based question-answering system built over an EU-focused fake-news dataset, with a data-integrity layer on top that guards the retrieval pipeline against tampering and data poisoning. Ongoing research, done with two PhD collaborators; a paper is in progress.
Key Features:
- Records each source document's identifiers (hash, filename, size) at upload time.
- Anchors those fingerprints on Polygon, with the underlying content stored via IPFS/Pinata.
- At inference time, re-verifies retrieved context against its recorded fingerprint before it ever reaches the model.
- Rejects context whose fingerprint no longer matches, catching tampering or data-poisoning attempts mid-pipeline.
Research Angle:
Explores how cheap, verifiable on-chain anchoring can harden RAG pipelines against a source-integrity failure mode that similarity search alone can't catch — a document can be semantically relevant and still have been silently altered after ingestion.