Projects

The through-line is agents that have to survive contact with a real system — a Kubernetes cluster, a live interview, someone's supply chain. Everything below was built end to end rather than as a notebook demo.

Kubeflow docs-agent

2026

Agentic RAG over the whole Kubeflow surface — docs, GitHub issues, manifests and source — served as an MCP toolset behind the Kubeflow website.

My Google Summer of Code 2026 project, and the thing I spend most of my open source time on. The starting point was a documentation chatbot that could only read Markdown. A Kubeflow user with a broken pipeline does not have a documentation problem, they have a "this exact error appeared in an issue eighteen months ago" problem, so the agent needed to read more than prose.

I added ingestion pipelines for GitHub issues, application code and Kubernetes manifests, and exposed them as separate MCP tools so the agent can choose where to look instead of retrieving from one undifferentiated blob. Every chunk carries path, product area, version and source metadata, which is what makes filtered retrieval possible at query time.

The largest change was merged as a single 5.4k-line PR covering a three-tool MCP server, TEI embeddings, the issues and code pipelines, and CI/CD onto OKE. Since then the work has moved to the edge: rate limits, CORS lockdown and anonymous session-JWT auth for the public chatbot, with the Istio configuration migrated into a Helm chart so the guardrails ship with the deployment rather than living in someone's cluster.

Python · MCP · Agentic RAG · Kubeflow Pipelines · KServe · Istio · Helm · OKE

CrackRound

2025

An agentic mock-interview platform: five streaming interviewer personas, a real-time voice loop, and a live code judge wired into the model context.

Built end to end, from an empty repository to something people pay per session to use. The interviewer is not a chat window with a prompt — it is a persona that holds a rubric, drives the round, interrupts, and follows up on a weak answer.

The voice loop is the part I am most happy with: streaming speech-to-text into GPT-4o and back out through streamed TTS over a WebSocket, at roughly 1.5 seconds end to end. Anything slower and the conversation stops feeling like an interview.

A DSA judge and a system-design whiteboard both feed into the model context, so the interviewer can see what you actually wrote rather than what you claimed. Scoring is JSON-schema constrained across five dimensions, with latency tracing and a hard $2-per-session cost ceiling.

Next.js · TypeScript · GPT-4o · WebSockets · Prisma · PostgreSQL · Sarvam

AirCab

2025

A voice-first booking agent that turns a spoken request into a confirmed ride, with the whole conversation under one tool-calling loop.

An experiment in how far you can push a voice agent when the task has real side effects. Booking a ride is a good test case because it is short, has a clear success condition, and punishes a model that hallucinates a confirmation.

The interesting problem was not the speech pipeline but the state machine underneath it: what the agent is allowed to assume, when it must read back a detail before committing, and how to recover when the user changes their mind three turns in.

Python · LLM tool calling · STT/TTS · FastAPI

FlowForge

2025

A node-based canvas for composing LLM workflows, where the graph you draw is the execution plan the runtime actually walks.

Visual agent builders usually stop at demo quality because the canvas and the runtime drift apart. FlowForge keeps one representation: the graph is compiled straight into the execution plan, so what runs is what you drew.

Nodes cover the usual set — prompts, tools, retrievers, branches, loops — and the runtime handles fan-out and joins. It exists because I got tired of rewriting the same orchestration glue by hand for every new agent idea.

TypeScript · React · ReactFlow · LangGraph · FastAPI

AI Learn

2026

Voice-to-voice learning platform that teaches interview topics in Hinglish, built as a Next.js PWA over a FastAPI provider layer.

Most tutoring products assume you want to read. This one assumes you are commuting. The teaching style is deliberately Hinglish because that is how the explanation actually sounds when a senior engineer walks a junior through a concept in an Indian office.

STT, TTS and the LLM sit behind swappable adapters, so the same session logic runs against mocks in tests and Sarvam in production.

Next.js · FastAPI · Sarvam · PWA

Smaller things

VizCode

2025

Generates step-by-step algorithm visualisations from a prompt, with Gemini emitting a structured trace the renderer replays.

Blog to Podcast

2025

Scrapes any public blog post and returns a listenable episode: Firecrawl for extraction, GPT-4 for the script, ElevenLabs for the voice.

Online handwriting recognition

2025

BiLSTM + CTC over pen-stroke sequences from IAM-OnDB — recognising handwriting from how it was written, not how it looks.

CodeNITW

2023

Placement-prep site for NIT Warangal students, with a Codeforces-API leaderboard that a few hundred people actually used.