aditya.com
--:--:--

Builder, CTO, Founder

Most stacks are rented. Mine, I can rebuild from zero.

I buildthe layerunderneath

Building01

tunyt.com — the agentic OS for events

Shipping02

Veris — a camera that proves its photos are real

Engineering03

Agentic backends at Pokus, Bengaluru

Present tense

Now

Four things are true right now: a company I'm CTO of, a camera I'm shipping, a backend I'm paid to run, a degree I'm finishing. That's the surface. From here, the page goes down one layer at a time.

Building

tunyt.com — the agentic OS for events

Shipping

Veris — a camera that proves its photos are real

Engineering

Agentic backends at Pokus, Bengaluru

Studying

B.Tech CSE (AI & ML), LPU — class of 2028

The surface

Products

The top layer the part of the stack with real users and money attached. Shipped, paid for, and still running while Im asleep.

Multi-agent orchestration across venue booking, ticketing, vendor coordination, and run-of-show — the whole operation an organiser would otherwise run out of a spreadsheet. Built from the ground up: agent workflows, real-time infrastructure, and the UI on top. I lead engineering direction, technical hiring, and system architecture. Tixin owned the ticketing primitive; tunyt owns everything that happens around it.

  • TypeScript
  • Agentic AI
  • Multi-agent workflows
  • PostgreSQL
  • Redis
~2M
lines of code

written across the full stack as CTO, agents through UI

Visit ↗private repo

Event-driven throughout: Kafka on the write path, Redis for read caching, CQRS to keep booking reads fast while writes stay consistent under burst. The whole thing auto-scales on a Kubernetes cluster I stood up by hand on two Raspberry Pi 5s, which is a deeply unreasonable place to serve a live on-sale from, and it held.

  • TypeScript
  • Kafka
  • Kubernetes
  • PostgreSQL
  • MongoDB
  • Redis
50K+/sec
events ingested

sustained peak through the Kafka pipeline

16K+/sec
requests served

on a two-node Raspberry Pi 5 Kubernetes cluster

−70%
query latency

after PostgreSQL and MongoDB tuning on the booking read path

−60%
deploy time

after the CI/CD automation rebuild

Visit ↗private repo

Microservices, distributed systems, and DevOps, taught to people one or two years behind me. The curriculum is the stack I actually run in production, built around hands-on labs for scaling systems, debugging live incidents, and real-world failure simulations rather than tidy exercises.

  • Node.js
  • Docker
  • Kubernetes
  • CI/CD
100+
students mentored

cumulative across backend and DevOps cohorts

The layer underneath

Primitives

When a primitive is load-bearing, I rebuild it rather than install it a load balancer instead of Nginx, a queue instead of BullMQ, a key-value store instead of Redis. The library is rarely the worse choice. Its that Id rather not run something on the critical path I can only reason about from the outside.

Speaks RESP well enough that redis-cli connects and doesn't notice. Writing the protocol parser, the storage layer, and the command handling by hand — no shortcuts — is the fastest way I know to understand why Redis is shaped the way it is.

  • Go

Built to understand what a load balancer actually does when it distributes a request. Dynamic auto-scaling for Bun backends, with active health checking, intelligent routing, and a real-time monitoring surface. Rebuilding it taught me more about connection handling than a year of configuring someone else's.

  • Bun
  • TypeScript

Command Query Responsibility Segregation with fully separated command and query paths, pulled out of the production system and made standalone so the pattern is legible on its own. This is the reference implementation behind Tixin's booking path.

  • TypeScript
  • Kafka
  • PostgreSQL

Account, catalog, and order services talking over gRPC, fronted by a single GraphQL gateway, backed by PostgreSQL and Elasticsearch with real-time WebSocket wiring. A complete microservice topology small enough to hold in your head and real enough to break.

  • Go
  • gRPC
  • GraphQL
  • PostgreSQL
  • Elasticsearch

Rooms, producers and consumers, and real-time media routing built on mediasoup. Multi-party video is where the naive mesh topology falls over, and an SFU is the thing you have to build to find out why.

  • TypeScript
  • WebRTC
  • mediasoup

Down to the metal

Edge & Hardware

The software runs out, and there is still a layer: hardware, inference that never leaves the device, integrity that survives the trip off it. The combination is the point a sensor that decides locally, and can still prove what it saw.

The camera derives an ed25519 identity from the device itself, signs every photo at the moment of capture, and a Solana Anchor program verifies that signature on-chain before anything is minted. Filecoin for storage, QR claim flow, proof of attendance. Built for a world where 'is this real' has stopped being answerable by looking — this kills deepfakes at the source rather than trying to detect them after.

  • Rust
  • Anchor
  • Solana
  • Filecoin
  • Embedded

A custom multimodal emotion model fusing vision, audio, and context, running entirely locally — no cloud API in the loop. It remembers you across sessions and adapts to how it can actually help. Latency and privacy both improve when nothing leaves the device. This is where the infrastructure work and the softer thread meet.

  • Raspberry Pi 5
  • ESP32
  • Python
  • PyTorch
  • Edge AI
private repo

A co-evolving multi-agent RL environment for cognitive security: the defender works a real intrusion while an adversary gaslights it with fabricated SIEM entries in real time. GRPO group sampling, adaptive curriculum, LLM-as-judge scoring. Built for the Meta × PyTorch × Hugging Face OpenEnv Hackathon India 2026.

  • Python
  • PyTorch
  • MARL
  • OpenEnv

Syncs with my tools, calendar, code, and agent setup, and has enough access to actually act rather than suggest. The point isn't a chatbot — it's that the cost of leaving flow to go find something is the tax I most wanted to stop paying.

  • TypeScript
  • LLM orchestration
  • Agent workflows

A BERT-style bidirectional transformer that generates text by progressively unmasking tokens from a fully masked sequence, rather than left-to-right. Trained on TinyStories locally on Apple Silicon — small enough to actually finish, real enough to show why diffusion over discrete tokens is awkward and interesting.

  • Python
  • PyTorch

An Anchor program holding escrow and settling payouts on-chain with no operator in the middle to trust. The interesting constraint is that every failure mode has to be handled in the contract, because there is no server to patch afterwards.

  • Rust
  • Anchor
  • Solana

A microservice platform that reads GitHub, LinkedIn, and LeetCode to level a student's actual skills, generates a personalised roadmap, and runs live AI-agent mock interviews with facial emotion detection to score how they hold up under pressure.

  • TypeScript
  • Microservices
  • AI agents
  • Computer vision
05Thesis
Everyconvenienceis rent

The honest map

Stack

A thesis like that obligates an honest inventory. A badge wall tells you what I've installed, not what I'd stake an incident on so this is graded instead: three rings, each with its honest qualifier attached, including the one that says I'm still learning.

Production

I ship this under real load, and I'm on call when it breaks.

  • TypeScript
  • Node.js
  • Kafka
  • Redis
  • PostgreSQL
  • MongoDB
  • Docker
  • Kubernetes
  • CQRS
  • Event-driven microservices

Working proficiency

Demonstrated and deployed, but less publicly load-tested.

  • Go
  • Python
  • gRPC
  • GraphQL
  • WebSockets
  • Next.js
  • Terraform
  • Ansible
  • Jenkins
  • GitOps
  • Elasticsearch
  • Event sourcing
  • Saga pattern

Exploring

Real projects, real commits, not yet battle-tested. Listed here rather than claimed above.

  • Rust
  • Anchor / Solana
  • C++
  • Embedded (RPi 5, ESP32)
  • Edge AI
  • Multimodal emotion models
  • Diffusion language models
  • Multi-agent RL

Around the work

Community

Sovereignty isn't solitude. The same work, proven in rooms competitions entered and won, classrooms taught, one hotel actually run.

Competitor

Brand Wars at IIT Mandi. Innovation Deck. BHU. Undefeated.

11 entered, 11 won

Instructor

Backend & DevOps at Learner's Arc

100+ students

Operator

Ran a hotel. Rebuilt how it sells.

+73% revenue in 3 months

Mentor

GirlScript Summer of Code

recognised for impact

Judge

Hackathons across the LPU-adjacent circuit

4+ judged

Speaker

Guest lectures on system design and backend engineering

Contact

Say hello back.

Let's build something that stays up: CTO and founding-team roles, hardware plus AI, research collaboration, or speaking. Also anything that has no business running on a Raspberry Pi. I read everything.

Aditya Dutt Pandey

Aditya Dutt Pandey