Jyotish Kumar

A single-instrument limit order book and TCP matching engine designed around deterministic matching and a hand-rolled, non-blocking epoll event loop.

• Implemented price-time priority matching with partial fills, cancellation, ownership checks, self-trade handling, and multiple-price-level sweeps.
• Served clients through a custom non-blocking epoll-based event loop rather than a thread-per-connection or networking framework.
• Sustained 64,500 concurrent TCP connections for a 2-minute load test with zero failures, at roughly 164K requests/sec aggregate throughput.
• Kept steady-state median latency around 42–45 µs while scaling the order book from 100 to 10,000,000 orders.
• Passed 105 tests under ASan/UBSan and TSan, plus dedicated TCP chaos tests covering drip feeds, fragmentation, and malformed oversized payloads.

Benchmarks were measured locally on an i5-11300H running Fedora Linux; detailed conditions and raw results are documented in the repository.

A single-node concurrent key-value store with a custom TCP protocol, command parser, Append-Only File persistence, and a full build, test, sanitizer, soak, and benchmark pipeline.

• Built a thread-safe in-memory store with Append-Only File persistence so writes can be reconstructed across restarts.
• Reached 218,251 requests/sec in high-stress benchmarking with 100 persistent concurrent threads.
• Completed a 10-minute endurance run processing 108.5 million requests with zero failures.
• Exercised the server at up to 64.5 concurrent client connections while keeping the implementation sanitizer-clean.
• Tested protocol resilience against fragmented and sticky TCP packets and validated 10,000-key data integrity with zero read/write errors.

Benchmarks were measured locally on an i5-11300H running Fedora Linux; exact test conditions and raw logs are available in the repository.