I'm an undergrad at BIT Mesra studying Mathematics and Computing, currently building hybrid search and retrieval systems and contributing to open-source LLM infrastructure. My work sits at the overlap of classical IR (BM25, ranking theory), dense retrieval, and applied ML systems.
- 🌱 Deep in Information Retrieval, RAG systems
- 🛠️ Active contributor to Haystack - deepset's production LLM orchestration framework
- 📐 Math background (Linear Algebra, Graph Theory) directly informs how I approach retrieval and ranking problems.
Contributing to a production Python framework for LLM orchestration and RAG.
| PR | Description |
|---|---|
#11581 |
Added run_async to LLMEvaluator, FaithfulnessEvaluator, and ContextRelevanceEvaluator - enables concurrent evaluation in async apps, with runtime detection of async-capable chat generators and automatic thread-pool fallback for sync-only ones |
#11552 |
Fixed a silent serialization bug in LLMEvaluator.to_dict that caused pipeline reloads to silently reset user-configured parameters |
#3412 |
Extended GoogleGenAITextEmbedder, GoogleGenAIDocumentEmbedder, and GoogleGenAIMultimodalDocumentEmbedder with timeout/retry support, full serialization round-trips, and test coverage |
Python rank-bm25 ChromaDB sentence-transformers NetworkX
A three-signal hybrid retrieval CLI fusing BM25 lexical search, dense vector search, and AST-derived call-graph structural retrieval via Reciprocal Rank Fusion.
- Evaluated across five retrieval configurations on a 30-query benchmark over ~1,260 indexed functions across scikit-learn's
model_selection,preprocessing, and other subpackages - MRR@5: 0.6022 and Recall@20: 0.7889 (BM25 + semantic + graph fusion) vs. MRR@5: 0.5122, Recall@20: 0.7222 semantic-alone, and MRR@5: 0.5389, Recall@20: 0.7389 BM25+semantic without the graph
- Implemented call-graph structural retrieval, scoring functions by call relationships to top BM25/vector results - summing decayed contributions (0.5^hop) from every reaching seed, weighted by inverse seed rank
- Identified degraded structural performance because of nodes colliding on bare function names (e.g. sklearn's repeated
fit/transformacross estimator classes), fixed via collision-free graph construction keying nodes on globally-unambiguous (name, class) pairs - driving the best overall fusion result (+0.063 MRR@5, +0.05 Recall@20 vs. BM25+semantic alone) - Shipped as a production CLI: OS-native keyring for API keys, SHA-256 multi-codebase index management, mtime-based staleness detection, lazy imports cutting startup latency from ~20s to ~6s
Python FastAPI ChromaDB Gemini API Streamlit
A modular AI backend for embedding-based recommendation over scraped product data.
- Automated scraping pipeline (BeautifulSoup) collecting 370+ product profiles
- Two-stage pipeline: candidate generation → LLM-based reranking, with fault-tolerant fallback across rate limits and provider outages
- Streamlit dashboard for real-time inspection of retrieved recommendations
Languages: 
ML / IR: 
Backend / Infra: 
Tools: 
Birla Institute of Technology, Mesra - B.Sc. (Hons.) Mathematics and Computing
Relevant coursework: Object-Oriented Programming (Java), Data Structures & Algorithms, Discrete Mathematics & Graph Theory, Information Retrieval