scifact

VerifAI initiative to build open-source easy-to-deploy generative question-answering engine that can reference and verify answers for correctness (using posteriori model)

🩺🔍 Combating Biomedical Misinformation through Multi-modal Claim Detection and Evidence-based Verification

A genral RAG Search chatbot, with SoTA RAG techniques such as HyDE, Hybrid retrieval with BM25 + RRF and Cross encoder reranking. Evaluated on the BEIR scifact dataset and compared all the different pipelines i tried along the way

Public release of SciFCheX system developed for COM3610 Dissertation Project at the University of Sheffield. The pipeline is designed to perform fact-checking on scientific claims.

Domain adaption for an embedding model using unsupervised and supervised finetuning on scientific texts for the SciFact retrieval task.

🩺⚡ SynthMed: Generating and Detecting Multimodal Deepfakes for Healthcare Communication

Evaluating retrieval strategy bias in biomedical scientific claim verification using RAG — comparing dense, BM25, hybrid, and query-reformulation retrieval across Support Recall, Contradiction Recall, and Balance Score on SciFact.

RAG evaluation on BEIR SciFact: BM25, dense and hybrid retrieval with LLM answers.

Verifies that each citation in a draft actually supports the claim it's attached to. Faithfulness, not existence.

Reproducible SciFact evidence-retrieval benchmark for scientific RAG

Evidence-first SciFact retrieval audit with source-aware splits, calibration, abstention, citations, and reproducible metrics.

SciFact retrieval experiments with BM25, hard negatives, custom InfoNCE dual encoders, and reranking.

Production-style RAG answering API with hybrid retrieval, grounded generation, citations, validation, caching, and tracing.

Chiral Narrative Synthesis workspace for Thinker/Tinker LoRA pipelines, semantic fact-checking, telemetry, and reviewer-ready CNS docs.

A claim-level fact-checking system that uses RoBERTa to spot unsupported or contradictory scientific claims.

Decomposing two-stage retrieval failure on SciFact into candidate-set vs. reranking errors, with a calibrator that predicts which queries will fail.

Browser-side IR benchmark: BM25 vs Semantic vs Hybrid retrieval on SciFact (BEIR). Bauman MSTU NIR 2026.

Production-style benchmark for RAG retrieval: dense, BM25, hybrid, RRF, metadata filters & query rewriting on BEIR SciFact. Docker, Qdrant, FastAPI, GPU embeddings.

Dataset management library for ML experiments—loaders for SciFact, FEVER, GSM8K, HumanEval, MMLU, TruthfulQA, HellaSwag; git-like versioning with lineage tracking; transformation pipelines; quality validation with schema checks and duplicate detection; GenStage streaming for large datasets. Built for reproducible AI research.

MSc dissertation: empirical analysis of retrieval effects and failure behaviour in retrieval-augmented scientific claim verification, with stance-aware reranking.