Long-form write-ups of systems I built from scratch: a distributed key-value database in C++, a URL shortener with multi-layer caching, custom memory allocators, and my published machine learning research. Each post shows the real code, the measured numbers, and what I would change next.
AtomicKV, part 1 of 2
How one AtomicKV node works: a single-threaded Linux epoll event loop, an O(1) LRU cache behind a Bloom filter, and an on-disk B-Tree in C++17, with benchmarks.
Read articleAtomicKV, part 2 of 2
AtomicKV as a masterless cluster: consistent hashing with virtual nodes, gossip failure detection, Lamport clocks, replication, read repair and anti-entropy, in C++.
Read articleShortify's design: two cache layers over PostgreSQL, a Redis Bloom filter, cache-stampede locks, Snowflake IDs in Base62 and a token-bucket rate limiter in Node.js.
Read articleFour C++ memory allocators written from scratch (linear, stack, pool, free list), how a limit order book uses them, and a fair benchmark against glibc new/delete.
Read articleOur ICDAM 2025 paper: sandbox behaviour of 4,280 Android apps turned into features and images, then classified by a Random Forest, CNNs and a ViT (99.78% accuracy).
Read articleAn offline ESP32 battery monitor from our IEEE paper: a physics-informed neural network (R² 99.70%) compressed to INT8 with Edge Impulse EON, plus no-start prediction.
Read articleHow Chatur answers questions about PDFs with RAG: PyPDF2 extraction, recursive chunking, Gemini embeddings in FAISS, a grounded prompt, and lessons on chunk size.
Read article