I work on retrieval systems and vector databases, mostly the layer where query processing, index behaviour, and search efficiency meet. I'm interested in how retrieval infrastructure holds up under real conditions: changing data, predicate filters, and the demands of downstream LLM systems, rather than just how it performs on a fixed benchmark.

I'm currently a B.Tech student in Information Technology & Mathematical Innovation at the University of Delhi, working independently on a small portfolio of research projects in vector search and retrieval, alongside coursework in algorithms, ML, and systems.

Ongoing A C++ evaluation framework studying whether query-aware ANN search-effort routing, calibrated on a static HNSW index, remains reliable once the index evolves through insertions and deletions. Includes a fully audited negative result for a static-feature router, and an ongoing extension testing policy validity directly under controlled index evolution.

Ongoing A C++ filtered-ANN benchmark (pre-filter, post-filter, ACORN) studying how predicate selectivity and local vector structure determine the search effort needed for target recall. Includes an independently found and audited cost-accounting issue in ACORN's native effort metric, and a structural decomposition of filtered-search recall into unreachable vs. under-searched targets.

Independent research project A controlled long-context evaluation varying context size, evidence position, and semantic similarity of distractors, to characterize how individual queries — not just aggregate averages — respond as retrieved context grows.

GSoC 2026, PEcAn Organization — contributing to an LLM-based extraction system that converts unstructured scientific text into structured, provenance-linked records with field-level confidence estimates. Ongoing, under review, not yet merged.

A few earlier projects from before I moved toward retrieval/ANN research:

• ML Deployment Framework — a small Dockerized API for serving ML models, built to get hands-on with reproducible deployment rather than just training models in a notebook.
• Email Classification Pipeline — a text-classification pipeline covering ingestion, preprocessing, and training/evaluation on a structured dataset.

Languages

• C++
• Python
• SQL
• Java

Retrieval / ANN

• HNSW
• FAISS
• ACORN
• hnswlib
• BM25
• Dense & hybrid retrieval

ML & Evaluation

• scikit-learn
• Statistical testing
• Experimental design
• Error analysis

Systems

• CMake
• Linux
• Git
• Docker
• CI/CD