From Rust model gateways and agentic QA to sharded blockchain consensus and early-college hacks.
Rust/Tokio-based model gateway that isolates streaming and the agentic loop from the data-plane API tier. Led the migration of OCI Generative AI’s OpenAI-compatible agent execution onto it — 15% lower p90 latency and 90% less memory.
A general QA agent built on the platform’s own agent framework that automates end-to-end regression testing across diverse scenarios and edge cases, speeding up releases across GenAI services.
Implemented the AHL (Attested HyperLedger) sharding protocol in ResilientDB, an open-source permissioned blockchain (later an Apache Incubator project). Enhanced PBFT with a two-phase-commit optimized protocol, increasing throughput ~3× to 240K txns/sec.
Implemented the BFT-SMaRt protocol inside ResilientDB, with major changes to its Paxos-based VP consensus module — improving throughput by 10%.
Automates grading of 190+ students’ C++ assignment submissions collected through the handin command, cutting manual grading effort by 85%.
A convolutional deep Q-learner that learns Atari Pong with 98% accuracy, plus representation learning to understand how the agent predicts moves.
A multi-class Support Vector Machine that recognizes digits with 85% accuracy.
Movie recommenders using spectral clustering and matrix completion, reaching 91% accuracy.
A framework that performs static and dynamic analysis on Android apps and produces an easy-to-understand safety rating.
Hosted PICT’s National Computing Contest: a sandboxed Python judge that compiled and ran submissions against hidden test cases and returned verdicts (AC, WA, CTE, RTE).
Shows library book availability, issued books and due dates, backed by a Node.js REST server over MongoDB.