Why a queue between ingest and delivery?
Notify commits the event first, then publishes to RabbitMQ, so a provider outage never blocks the caller and every job can be retried or recovered from the database.
I am an AI engineer and full-stack developer, in my second year of a B.Tech in Computer Science (AI) at Vedam School of Technology, Pune. I build agents and RAG systems, ship the backend, web and mobile around them, and evaluate LLMs for a living.
I started college solving LeetCode problems. Four months later I was building multi-agent systems at hackathons, and by the summer I was the only engineer on a client's backend, PWA and app. Right now I do three things at once: full-stack intern at Third Shade Media, certified LLM evaluation expert at Deccan AI Experts, and co-founder of CampusCritique.
I was born in Jamshedpur, my family is from Bihar, I grew up and went to school in Delhi, and I study and work in Pune. School also left me with some German, a Scouts badge (Dwitiya Sopan) and a year as a prefect. I watch a lot of anime; Bleach: Thousand-Year Blood War is the current one.
The things I care about are simple: build for a specific person, measure what shipped, and say plainly what did not work. Every number on this site links to its source, and the failures get the same space as the wins.
Notify commits the event first, then publishes to RabbitMQ, so a provider outage never blocks the caller and every job can be retried or recovered from the database.
KisanMind filters and scores with rules first and lets the model rerank the shortlist. Deterministic where it can be, testable, and cheaper to run for a farmer on a phone.
OilTrace runs on patches where oil pixels are rare; Dice loss kept collapsing to the empty mask. Focal loss concentrates the gradient on the pixels the model gets wrong.
Payment gateways retry, reorder, and send refunds through the same pipe. CampusCritique keys on the gateway event id and guards refund updates.
Margo Rubber had contradictory source material. A build that fails on a missing fact is cheaper than an apology to a customer.
Ask this site answers only from retrieved chunks above a calibrated similarity threshold and says so when nothing qualifies. A wrong answer about me costs more than no answer.
8 projects on Work. Every one is sourced; the ones marked measured have an evaluation table.