AI product · Co-founder
Parivar AI
A Hindi and Hinglish family companion. I built the backend for voice calls, persona conversations, and long-term memory.
Node.js · PostgreSQL · pgvector · LLM APIs
Google Play · Currently being redeployed
I’m Krrish Kumar. I build Node.js APIs, background jobs, and real-time features. At Parivar AI, I built the backend that lets you call an AI persona in Hindi and pick up a conversation later.

I built the backend for each of these projects. Pick one to see what happens after the user taps a button.
AI product · Co-founder
A Hindi and Hinglish family companion. I built the backend for voice calls, persona conversations, and long-term memory.
Node.js · PostgreSQL · pgvector · LLM APIs
Google Play · Currently being redeployed
Real-time backend
I built a ride-matching service that finds drivers near a pickup with Redis and sends each one an offer over Socket.IO.
Redis GEO · Socket.IO · MongoDB
Microservices
An airline booking backend split into five small services, so a slow email server never slows down a booking.
RabbitMQ · API gateway · MySQL
Code execution platform
I built a code judge that queues submissions in BullMQ, runs each one in a disposable Docker container, and sends the verdict back to the browser.
BullMQ · Docker · Fastify
Pick a project, then step through one request from the user’s first action to the result.
Which drivers should see your request, and what happens to the other offers when someone accepts?
Read the RideWave case studyYou choose a pickup and destination. The API estimates the fare from the distance and saves a pending booking.
What happens between speaking into your phone and hearing the AI persona answer in Hindi?
Read the Parivar AI case studyYou start a voice call with Dadi in the Android app. WebRTC carries your microphone audio to a LiveKit session.
Which parts of a flight booking have to finish before the passenger gets a confirmation?
Read the Airline Booking System case studyThe gateway checks the request limit and asks Auth to verify the passenger’s token before forwarding the booking.
How does a code submission become a verdict in the browser while the API stays free to handle other users?
Read the AlgoForge case studyThe API saves your solution as pending, fetches the problem’s test cases, and returns a submission ID.
I publish small tools for problems I run into building backends. These are my two latest releases.
When several workers call the same model, each sees only part of the bill. llm-pacer gives them one Redis-backed budget for requests, tokens, and concurrent calls.
An offline check for credentials exposed to npm install scripts. It lists what needs attention and suggests a rotation order while keeping secret values out of the report.
Send the same webhook twice and check whether the customer was credited twice. It also tests deliveries that arrive late or out of order.
Run the webhook test suite from an editor through MCP and compare the results after changing a handler.
Catches webhook handlers that parse a body before checking its signature or process a delivery twice.
Flags mistakes in BullMQ workers, including missing error listeners and job IDs that fail to prevent duplicate work.
Starts a TypeScript queue service with an API, a worker, health checks, and graceful shutdown already wired up.
Validates Indian identifiers in Zod, including GSTIN, PAN, IFSC, and UPI IDs. It also formats rupee amounts in lakhs and crores.
Dec 2025 – Present
Launched on Google Play, being redeployed
Co-founder & Backend/AI Engineer
Oct 2025 – Dec 2025
Backend Developer Intern
2022 – 2026
Ajay Kumar Garg Engineering College, Ghaziabad, India
Three calls I made in my own projects, and what each one cost.
I keep driver locations in Redis and send the booking to the sockets that match a radius search. I also keep track of who got it, so I can take it off their screens when someone accepts. The part I’d change next is confirmation. Two drivers can still accept at nearly the same time.
I keep the seat update in the booking request and put the reminder on RabbitMQ. A slow mail server can delay the receipt while the passenger already has a booking result. That leaves me with a separate reminder worker to run and monitor.
My co-founder built the Flutter app. I put the speech-to-text, model reply, and speech playback in a LiveKit agent, and kept conversation memory in the backend. That gave us a separate place to work on call quality, at the cost of another service to deploy.
Voice and memory are already part of my production work. I’m now building retrieval pipelines and learning how to evaluate them.
Shipped
Parivar AI talks to users in Hindi, calls them in a natural voice, and reacts to the photos they send.
LLM APIs · Prompt design · Conversation management · Text-to-speech · Image input
Building
Retrieval over your own documents, and structured outputs other services can consume.
RAG · Embeddings · Vector search · Structured outputs · FastAPI · Tool calling
Next
Agents with bounded autonomy, measured by evaluations and held in place by guardrails.
Agents · LangGraph · MCP · Evaluations · Guardrails · LLMOps
LLM features behind real APIs: typed inputs, typed outputs, streamed responses.
Retrieval pipelines that ground answers in sources and can prove it.
Stateful, tool-using workflows with explicit control flow and bounded autonomy.
Speech and vision as first-class inputs and outputs.
The backend half of AI: evaluation, guardrails, observability, cost, and latency.
Enough of the model side to reason about what the API is doing.
Not claimed: ML research or training models at scale. The deep learning layer is there so I understand what the model is doing behind the API.
The tools I’ve shipped with and the ones I’ve used in hands-on projects.

I like figuring out what happens after someone taps a button.
I got into backend work building a flight booking system. Getting one passenger a seat and an email took five services. Later, on RideWave, I had a different problem to solve. Several drivers could see the same offer, but only one should get the ride.
In December 2025 I co-founded Parivar AI. My co-founder built the Android app, and I built the service behind the conversations. These days I’m interested in how to tell whether an AI remembered the right detail from an earlier chat.
Try projects to find a case study, packages for my npm releases, or prs to read a change that made it upstream.
I'm open to backend and AI engineering roles, founding-engineer opportunities, and freelance work.
whoami
Krrish Kumar · Backend & AI Engineer · Delhi NCR, India
I build APIs and background workers in Node.js. I co-founded Parivar AI and built its voice and memory backend.