Varuni H K

Research

Labs and groups I'm working with.

• Cornell Database Group

  with Prof. Immanuel Trummer

  Quantum-inspired distributed index replica placement.

• UIUC REU Program

  with Prof. Tianyin Xu

  Simulating real-world faults in SREGym.

• MIT NANDA

  with Prof. Ramesh Raskar

  NandaTown, and a distributed agent registry that holds up under failures.

• MIT

  with Tom Pollard

  Contributing to Croissant, a metadata format for ML-ready datasets.

• RISE

  research mentee with Dr. Seyyedeh Qazale Mirsharif

  Longitudinal embedding delta analysis for lung cancer CT scans.

Publications

2026

1. Croissant Baker: Metadata Generation for Discoverable, Governable, and Reusable ML Datasets

   NeurIPS 2026, Datasets Track

   arXiv

2. Co-Need Is Not Similarity: Locality Axes for LLM Agent Memory Retrieval

   NeurIPS 2026, MLSys Workshop

3. Keli: A Self-Paced Adaptive Platform for Oral-Aural Learning as a Sociotechnical Equalizer for Hearing-Impaired Students in India

   NeurIPS 2026, AI4GOOD Workshop poster

   NeurIPS page

4. Identity-Corrected Embedding Deltas for Lung-Nodule Progression

   NeurIPS 2026, Women in Machine Learning (WiML) sole author

   NeurIPS page

5. Gamifying Reality Orientation for Personalized Dementia Support for Senior Citizens

   ACM womENcourage 2026

6. SubGINs: Approximate and Exact Subgraph Matching with GINs

   ICANN 2026, Neural Networks for Graphs and Beyond lead mentor

7. Efficient Indexing for Google Zanzibar Based Authorization Systems Using Graph Dynamic-Transitive Closures

   28th International Conference on Advanced Communications Technology (ICACT), Pyeongchang lead author

8. Polaris: Multi-Agentic System for Conversational Enterprise Analytics

   AAAI-26 Workshop on Agentic AI Benchmarks and Applications for Enterprise Tasks, Singapore lead author

   PDF

2025

9. GAT: Grapheme Aware Tokenization for Kannada

   MIT Undergraduate Research Technology Conference (URTC) solo author

   GAT-K tokenizer Slides

Earlier

10. Exploring Caching and Indexing Mechanisms for Zanzibar Based Authorization Systems

    11th International Conference on Applied Systems Innovation, Kyoto

11. Leveraging Permissioned Blockchain for Securing Controlled Drug Prescriptions in India

    6th International Conference on Blockchain Computing and Applications, Dubai

Talks

• Dec 2025

  AI Today and Tomorrow track

  Grace Hopper Celebration India (GHCI)

• Sep 2025

  Maps, Masala & Missed Buses

  IndiaFOSS 2025, Data devroom

  Slides

  Got the audience to laugh, and I wear that badge with pride (with my silence of the lamps joke).

• Sep 2025

  Fast Authorization Checks with Dynamic Graph Indexing in Go

  GopherCon India, youngest speaker

  Slides

• 2025

  Grapheme Aware Tokenization

  MIT URTC

  Slides

• 2025

  Workshop on distributed graphs

  RootConf 2025

• Jun 2025

  Building an Incremental Index in Go: Concurrency Patterns Learnt and Performance Challenges

  Women Who Go Bangalore

  Slides

Projects

• semops

  Experimental semantic operators with Couchbase, built on GSI vector search to speed up retrieval over operational data.

• Transitive-closure caching and indexing for Zanzibar-like systems

  Dynamic transitive-closure indexing for authorization checks. Cut read latency by 57% over single-node SpiceDB on a 99% read workload. Go, Python, SpiceDB.

  Slides

• Litmus: a honeypot server for NandaTown

  Tests AI agent safety with trap layers such as exposed credentials, dangerous command lures and prompt injection, with tripwire links and signed scorecards. 3rd place at MIT NandaHack.

  Code

• Fuzzy and semantic alignment of multi-perspective narratives

  A search engine across Mahabharata retellings told from different characters' points of view, combining typo-tolerant keyword ranking with embedding-based semantic search.

  Write-up

• Project Roots

  A Flutter app of personalised cognitive games that cognitive specialists use for memory reinforcement with their patients. Funded with USD 6,500 by Cisco thingQbator.

  Code

Writing