AVNLP

Agentic Graph RAG for Medical diagnosis and Clinical QA. Combines LangGraph-orchestrated multi-hop reasoning with parallel graph, vector, community, and text-to-Cypher retrieval with an agentic plan–research–verify loop for evidence-grounded answers on MedQA, MedXpertQA, MedCaseReasoning and MMLU-Pro.

Advanced LLM fine-tuning techniques: SFT (LoRA, QLoRA, DoRA, P-/Prefix-Tuning), GRPO, DPO, ORPO, KTO & PPO; composable correctness/format rewards + LLM-as-a-Judge evals (DeepEval, Evidently AI) across math, multi-hop, medical & general QA on Llama 3, Mistral, Phi-4, Gemma & Qwen3. Built on TRL, PEFT & Unsloth.

Self-Reflective Question Answering for Biomedical Reasoning. GRPO fine-tuning via QLoRA & Unsloth with rewards for correctness, relevance, groundness, utility & XML structure. Structured think → answer → self-reflection with context grading, relevance assessment & groundness evaluation. DeepEval LLM-as-a-Judge (GEval, Faithfulness, Relevancy).

Training code for advanced RAG techniques - Adaptive-RAG, Corrective RAG, RQ-RAG, Self-RAG, Agentic RAG, and ReZero. Reproduces paper methodologies to fine-tune LLMs via SFT and GRPO for adaptive retrieval, corrective evaluation, query refinement, self-reflection, and agentic search behaviors.

Modular LLM ranking library for Information Retrieval and RAG. Implements state-of-the-art Pairwise, Setwise, and Listwise ranking with structured generation and specialized models (RankZephyr, RankLlama). Features efficient sorting algorithms, sliding windows, and zero-shot capabilities.

Production-ready Haystack/LangChain pipelines for Hybrid & Parent-Child Retrieval, Diversity Filtering, MMR, Metadata Filtering, Reranking, Query Enhancement, Multi-Tenancy, Agentic RAG across Pinecone, Weaviate, Milvus, Qdrant, Chroma with dataset loaders, evaluation tooling, and benchmarks for TriviaQA, ARC, PopQA, FactScore and Earnings Calls.

Advanced RAG pipelines for medical (HealthBench, MedCaseReasoning, MetaMedQA, PubMedQA) and financial (FinanceBench, Earnings Calls) QA. LangGraph orchestration + BAML structructed generation, Milvus Hybrid search (Dense + BM25 + RRF), three-layer Metadata Enrichment, Contextual AI instruction-following reranker, and DeepEval evaluation.

Pairwise Ranking Prompting (PRP): Zero-shot LLM reranking library implementing efficient pairwise strategies (Heapsort, Sliding Window, All-Pairs). Mitigates position bias via bidirectional comparison and ensures reliability with structured Pydantic validation. Built for Haystack pipelines.

Medical Reasoning Evaluation

Group Relative Policy Optimization (GRPO) implementations - NanoAhaMoment, GRPO:Zero, Simple GRPO, and GRPO from Scratch - spanning vLLM + DeepSpeed, custom Transformer stack, Bottle HTTP reference server, and pure PyTorch. Compares generation backends, reference policy strategies, reward designs, and loss functions on GSM8K and Countdown tasks.