Software engineer on Oracle AI Vector Search. I work on how vectors get compressed, indexed and searched inside the database.
Now
Oracle AI Vector Search
Before
AWS Redshift, Oracle India
Studied
Purdue (MS CS), IIT Delhi (B.Tech EE)
Published
NeurIPS 2025, ICONS 2020, BioCAS 2019
Press a key above to flip the board.
About
I'm a software engineer at Oracle in Redwood Shores, where I own the scalar-quantized IVF vector indexes behind Oracle AI Vector Search. I work on how vectors are compressed, built into indexes and rescored at query time, and I follow new quantization research like TurboQuant and RaBitQ to see what belongs in future IVF and HNSW indexes.
Before this I did my MS in Computer Science at Purdue, where I worked on reinforcement learning for LLM reasoning. I've also interned at AWS on Redshift, spent two years on storage features and backend services at Oracle in India, and studied electrical engineering at IIT Delhi, where my first paper was about training neural networks on analog hardware.
Jun 2025 – nowOracleRedwood Shores, CaliforniaSoftware EngineerNow+
Own scalar-quantized IVF vector indexes for Oracle AI Vector Search: about 75% less memory and 50% faster builds at the same recall.
Built query-time rescoring and a latch-protected quantization metadata cache shared by index build and query execution.
Prototype techniques from new quantization research (TurboQuant, RaBitQ) for future IVF and HNSW indexes.
Maintain the LOB storage engine, including a fix for a race between concurrent LOB access and a cross-session truncate.
May – Aug 2024Amazon Web ServicesSeattle, WashingtonSoftware Engineer InternDeparted+
Built an automated root-cause analysis tool for Redshift cluster failures during creation, deletion and patching.
Built a RAG pipeline that gives on-call engineers LLM-drafted resolution steps and triage information.
Oct 2022 – Aug 2023OracleBangalore, IndiaSoftware EngineerDeparted+
Built a sampling-based compression advisor for BasicFile and SecureFile LOBs that estimates savings without converting the full dataset, cutting analysis time by up to 80%.
Wrote a lex-yacc transpiler from Firebase Security Rules to SQL access-control rules, and designed the storage and enforcement layers of a backend-as-a-service product.
Jul 2021 – Oct 2022OracleHyderabad, IndiaSoftware EngineerDeparted+
Split the Fusion Payables monolith into microservices and defined their API contracts.
Built a service and admin dashboard for flagging and resolving payables exceptions.
Added usage and failure-rate metrics to a unified telemetry dashboard for load testing.
Research
Tap a row for details
Publications
YearVenueTitle
2025NeurIPSRevisiting Reinforcement Learning for LLM Reasoning from a Cross-Domain Perspective+
Z. Cheng, S. Hao, T. Liu, F. Zhou, Y. Xie, F. Yao, Y. Bian, N. Dey, et al. Datasets & Benchmarks Track. Introduces Guru, a 92K-example RL reasoning corpus across six domains. Guru on GitHub ↗
2020ICONSReduction of the Weight-Decay Rate of Volatile Memory Synapses in an Analog Hardware Neural Network for Accurate and Scalable On-Chip Learning+
ACM International Conference on Neuromorphic Systems.
2019BioCASOn-chip Learning in a Conventional Silicon MOSFET Based Analog Hardware Neural Network+
N. Dey, et al. IEEE Biomedical Circuits and Systems Conference. From my B.Tech thesis at IIT Delhi: backpropagation and on-chip training with MOSFET-based synapses.
Projects
DatesProject
2024 – 2025Teaching LLMs to reason with synthetic dataPurdue · LLM360+
Designed domain-specific reward functions and generated synthetic Zebra-puzzle data for Guru. Built a pipeline that generates morphology, world-model and natural-language proofs as training data, then post-trained models with PPO, GRPO and DPO.
Spring 2024Safer code generationPurdue+
Fine-tuned StarCoder with vulnerability detection as the RL signal, using a reward that blends CodeBLEU and BERTScore so security gains don't cost correctness.
Spring 2024Sharded key-value store on PaxosPurdue+
A linearizable, fault-tolerant key-value store with distributed transactions. A Paxos-backed shard master rebalances load by moving shards between replica groups.