Kanishka
Kumar

Software engineer at Skynift. I build the LLM pipelines, billing and auth that two AI products run on.

I am Kanishka, a software engineer at Skynift. I write backends and the AI pipelines that sit behind them: the LangGraph pipeline that turns a landing page into ad campaigns, the billing and auth under Helpexa, and the single-use sign-on tokens that let one product serve many brands. I like systems that still make sense on a second reading.

Away from the keyboard I reread Tolkien, keep up with Frieren, shoot street photographs, and defend a Clash of Clans village with more walls than it needs. This page is set the way I would want a book about all that to look: woodcut beasts, laid paper, and ink that changes with the hour.

The work

Four things I have built or am building, each drawn as the animal it most behaves like. Every claim here is one I can defend in an interview.

The three-headed hound. Three heads on one leash: three model providers answering to a single state machine.

Markitect

Skynift, Drac

A LangGraph pipeline that reads a landing page and writes a full ad campaign, with Claude, Gemini and OpenAI working inside one state machine.

What I did. I built the multi-model orchestration: the LangGraph state machine, the routing between providers, and the Zod-typed structured outputs every step hands to the next.

Crawls the landing page first and distills it into a brand profile before anything is generated.
One state machine runs Claude, Gemini and OpenAI image generation in sequence, each step typed.
Produces five kinds of ad content from a single URL.
Every model response is validated as structured output, so nothing downstream parses free text.

3 model providers
5 ad-content types
6+ pipeline stages

Made with

LangGraph
Claude
Gemini
OpenAI
Node.js
TypeScript
Zod

The hermit crab. It carries its whole house onto any shore, which is what a one-script-tag support desk does.

Helpexa

Skynift, Helpexa

An AI support chatbot a customer can put on any website with one script tag, built as three services: the platform, a serverless core, and the widget itself.

What I did. Top contributor on all three services. The serverless re-architecture was a team effort; I was its lead contributor.

Platform backend in FastAPI: JWT and cookie auth, OAuth, email verification, multi-tenant organisations and teams.
A Razorpay pay-as-you-go billing state machine that handles the whole subscription lifecycle.
Serverless core on AWS Fargate: Crawl4AI and Playwright scrape the knowledge base, n8n runs retrieval over OpenAI embeddings, quotas and rate limits keep it fair.
The widget is React compiled down to one widget.js, so a single script tag puts the chat bubble on any site.

#1 contributor on 3 production services
240+ commits across 8 repos

Made with

FastAPI
Python
AWS Fargate
Supabase
Razorpay
n8n
React
TypeScript

The griffin with a key. A gatekeeper that hands each visitor one key and burns it the moment it turns.

Drac whitelabel and SSO

Skynift, Drac

One product serving many brands: whitelabel multi-tenancy, with single sign-on tokens that can only ever be used once.

What I did. Primary author of the whitelabel and multi-tenancy subsystem on the backend, and owner of the theming engine on the client.

PS512-signed SSO tokens are consumed atomically when redeemed, so a replayed token finds nothing left to redeem.
A theming engine gives every tenant its own brand without its own codebase.
Super-admin roles and administration across organisations.
Firebase tenant provisioning and Backblaze B2 storage, kept separate per tenant.

Made with

Node.js
Express
React
RTK Query
Firebase
JWT (PS512)

The chameleon. Still being cut. A chameleon colours each patch of itself separately, which is the whole idea.

CineGrade AI

Personal project, in progress

Colour grading that treats a frame the way a colourist does: segment the scene, then grade each region on its own terms.

What I did. Solo project, in active development. The architecture is built and running. There are no users or metrics yet, so I am not claiming any.

SegFormer and SAM2 split each frame into regions: sky, skin, subject, background.
Each region gets its own grade instead of one global pass.
FastAPI and Celery run the jobs, GPU inference happens on Modal, and progress streams over a WebSocket.
Next.js front end with Clerk auth; finished renders live on Cloudflare R2.

4+ regions per frame

Made with

PyTorch
SegFormer
SAM2
Modal
FastAPI
Celery
Next.js
Cloudflare R2

What I work with

The tools, grouped by the kind of problem they solve.

AI systems

Pipelines that call models and get typed answers back.

Claude, OpenAI and Gemini APIs
LangGraph
LangChain
Retrieval (RAG)
Embeddings and vector search
Prompt design
Agent pipelines
n8n

Backend

The part that has to be right at three in the morning.

Python
FastAPI
Node.js and Express
TypeScript
Pydantic
REST
JWT and OAuth
WebSockets
Microservices
Multi-tenancy
Rate limiting
Celery

Frontend

Interfaces that feel simpler than what is underneath.

React
Next.js
RTK Query
Tailwind CSS
MUI
Vite
Motion

Cloud and data

Where it runs, and where it is kept.

AWS Lambda
AWS Fargate
API Gateway
S3
CloudFormation
Docker
GitHub Actions
Cloudflare
Firebase
PostgreSQL
Supabase
MongoDB

The road so far

Three milestones. The hare is where I am; the snail never left the first one.

BIT Mesra
B.Tech, Computer Science and Engineering
2020 to 2024
Four years of data structures, operating systems and late nights. Graduated with a CGPA of 8.02.
Explorex Technologies
Full-stack web developer intern
March to July 2024
First production work: Next.js, TypeScript, Tailwind and Framer Motion, the same family of tools this page still uses.
Skynift
Software developer
June 2025 to now
Intern to full time in four months. I work on Drac's AI marketing pipelines and Helpexa's support chatbot: billing, single sign-on, the serverless re-architecture, and the multi-model orchestration above.