Things I have built , shipped, and learned from.

A high-performance algorithmic trading framework for testing strategies on live Polymarket markets.

• Built a low-latency C++20 engine that processes live WebSocket feeds using lock-free SPSC ring buffers and stack-allocated containers on the critical path.
• Engineered a C++ / Python hybrid runtime with pybind11 embedding, enabling Python strategies to execute inside the C++ event loop with sub-millisecond callback overhead.
• Used PGO builds and heaptrack-driven profiling to remove per-message allocations in the hot path and drive p99 message-processing latency below one microsecond.

A high-performance inference engine optimized around asynchronous verification for continuous decoding.

• Architected an asynchronous speculative decoding engine that decouples draft token generation from target-model verification.
• Built a lock-free concurrent pipeline in Rust to overlap draft and target forward passes, increasing token throughput by roughly 50%.
• Kept the system deterministic while improving GPU occupancy and reducing memory-bandwidth bottlenecks during decoding.

An alternate memory allocator tuned for high-frequency allocation patterns.

• Achieved 2.1x average speedup over glibc malloc across 40+ benchmarks designed around real-world allocation patterns.
• Engineered per-thread arenas with slab-based segregated free lists for O(1) allocation and deallocation.
• Implemented a lock-free global page stack using CAS with tagged pointers and applied hugepages, compile-time dispatch, and hot-path inlining to keep allocator overhead low.

A full-stack application powered by a state-of-the-art agentic RAG (Retrieval Augmented Generation) pipeline, which answers the most difficult multi-hop quantitative & qualitative queries from information-heavy financial documents such as SEC 10-K reports.

A full-stack application made for HDFC Bank which lets them manage the transactions of their related parties by collating data from multiple sources and surfacing important trends in a single dashboard.

A ChatGPT equivalent for IIT Bombay that answers anything related to the institute — academics, hostels, fests, whatever you throw at it.

• Offered to over 14,000 students at IITB.
• Over 12,000 questions asked within three months.
• Answers using live information from internal systems and the open web.
• Quick response times — typically 6–7 seconds.

A desktop application that lets you control your computers from a single place.

• Turn on and shut down your machines.
• Remote into your machines from one dashboard.
• Store passwords for every machine in a single encrypted store.