I'm Rhythm. I build the backend of an enterprise AI agent platform: agent context management, RAG, tool and skill execution, secure connectors and OAuth credential lifecycles.
A product can look like one screen and one endpoint. Everything that makes that screen trustworthy lives underneath it.
Fetch data from a database, return it, show it. The user sees a list that loads. That is the whole promise, and it has to feel effortless.
Real engineering begins after the endpoint works. Add caching. Run jobs that keep the data pure. Put a queue in front of the slow path. Make it fast, correct and efficient.
Data rarely lives in one place. Different databases serve different shapes of question, and the system has to gather them into one consistent result.
All the user ever sees: data is coming.
Hot reads answered from memory. Invalidated when the truth changes, not on a hopeful timer.
Scheduled work that reconciles, refreshes and cleans, so stored data stays pure.
Kafka on the paths where work should be absorbed, ordered and retried instead of blocking the caller.
The right store for each question, stitched together so the caller gets one coherent answer.
Fast, correct and efficient, without the user ever having to know.
An AI-native platform for enterprises. Customers connect their business systems and build AI agents, workflows and apps without writing code.
Every one of these started from curiosity, a course, or a challenge. The early ones were written entirely by hand, before AI coding tools, and that is where I learned how things actually work.
Built to solve a problem I actually have, and to learn how agents really behave when they run every day.
Written line by line, by me. Debugging without a model to ask is how the fundamentals stuck.
A travel-planning site: explore destinations, build an itinerary, manage bookings. Motive: build a full product end to end across front end, PHP and MySQL.
Anomalies, health score and maintenance needs for CNC machines, shown on a live dashboard. Motive: take machine-learning models all the way to something an operator can use.
Scrapes a product's reviews and classifies each as human-written or AI-generated. Built for Winter of Code 7.0 by MSTC, DA-IICT. Motive: trust in online reviews.
Course allocation as an integer linear program: students rank choices, the solver maximises total preference within seat limits. Motive: use optimisation on a real campus problem.
Train search, fare lookup and booking flow. Built for Winter of Code 2024. Motive: learn a full MERN stack by rebuilding a service everyone knows.
A small desktop text editor, from the Winter of Code 5.0. Motive: learn how desktop apps handle events, menus and files.
A task-management REST API with CRUD, filtering, search and statistics. Motive: practise the Quarkus stack I now use at work.
UML design, code inspection, black-box and mutation testing from my fifth-semester course. Motive: learn how to prove software works, not just write it.
A competitive-programming and entrance-exam background that still shapes how I break down hard problems.
Happy to talk about platform engineering, AI agents, or a hard backend problem.