Hi, I'm Ayush — Software Engineer.

Embedded LSM-tree key-value store in Rust — WAL, memtable, block-based SSTables, size-tiered compaction, and Bloom filters. Benchmarked at 1.87M writes/sec (fsync off) and 24× faster absent-key lookups with Bloom filters vs without.

Async reverse proxy / API gateway in Rust (tokio, axum/hyper). Least-connections load balancing cut traffic to a slow backend from ~33% to ~2% vs round-robin. Sharded per-client token-bucket rate limiting at ~54ns/check, validated across 500 clients and 75,000+ requests with zero incorrect rejections. Benchmarked at ~70k req/s, ~5ms p99 — within ~1% of nginx at matched concurrency.

Autonomous GitHub issue-resolution agent: submit an issue, AI plans the fix, executes tool calls in an isolated Docker sandbox, and opens a PR end-to-end. Live SSE dashboard streams every LLM call and tool invocation. Hybrid LLM router across 5 providers with Redis-backed per-task $0.10 budget guard; shipped 8-file fixes in ~15 min and sub-$0.02 fixes in ~7 min.

MetaTrader-5 trading journal SaaS that auto-imports every fill, surfaces behavioral insights (tilt, revenge trades, session patterns), and delivers weekly AI reviews. Built the live AI trading coach with Groq over SSE, Redis rate limiting, and a Python/FastAPI MT5 bridge syncing broker trades in real time.

Production movie review and newsletter platform serving 36,000+ active users with 253,000+ monthly events. Sub-100ms API response via Redis caching, deployed on AWS with Docker CI/CD, SES, and BullMQ.

6-agent AI platform achieving 99.83% fault detection accuracy on the SWaT dataset (Random Forest), with a real-time WebSocket dashboard over 51 sensor points.

Adaptive JEE/NEET Chemistry prep platform with 25 chapters, 100 subtopics, and 325 concepts modeled on Neo4j with Ebbinghaus forgetting curve and IRT-based scoring.