Systems · Time · Power · Failure — consequences over intent
— Aniket Pal —
The Editor in Chief
Aniket Pal
Backend Engineer · Systems Thinker · Builder
I design backend systems and AI pipelines that hold up when
things get messy — APIs under load, RAG that retrieves the right
thing, agents that don't spiral. Self-taught, project-driven,
and more interested in what breaks than what looks good in a
demo. Everything here — the builds, the logs, the failures — is
the real working record, not a highlight reel.
A high-performance HTTP proxy written in modern C++ — connection pooling, an epoll event loop, and a config language small enough to memorize.
C++17 epoll Sockets CMake
▸ Full story
▾ Fold it back
[ project screenshot ]
DevProxy is an HTTP/1.1 forward and reverse proxy built from raw sockets up. It runs a single-threaded epoll-based event loop per core, keeps upstream connections alive in a slab-allocated pool, and reloads configuration on SIGHUP without dropping in-flight requests. Structured access logs ship as newline-delimited JSON. Built to understand exactly what nginx does for me — and what it costs.
A biomedical retrieval-augmented generation pipeline that answers clinical questions with citations, not confidence.
Python Ollama ChromaDB FastAPI
▸ Full story
▾ Fold it back
[ project screenshot ]
BNS RAG ingests biomedical literature, chunks it with section-aware splitting, embeds passages into ChromaDB, and serves grounded answers through a locally hosted Ollama model. A reranking stage filters retrieved passages before generation, and an evaluation harness scores every answer for citation faithfulness. The reranker cut hallucinated references by roughly 60% in testing.
We talk about flow as if it were weather — something that arrives, blesses an afternoon, and leaves without explanation. Four hours pass like one, the code pours out, and we credit inspiration.
