About

I'm a software engineer. I work at Quinn, where most of my time goes to pipelines that have to produce consistent output from models that would rather not — OCR extraction, quality control loops, and the scheduling and queueing around them. Before that I interned at Stoira and Frontend Hire, mostly on React Native products.

I did my B.Tech in Computer Science at KNIT Sultanpur, graduating in 2025 with a CGPA of 8.6. The interest started earlier than that, with a school computer teacher who explained algorithms as something you use rather than something you memorise.

What I build

Lately, backend and infrastructure. Seline is a Go platform that runs a personal AI agent in a container per user, which turns out to be mostly a lifecycle and scheduling problem rather than a model one. Radius is a Python MCP server that makes a personal bookmark corpus queryable — FTS5 and fuzzy search over libSQL, behind signed tokens with per-tool scopes. Playgroundish is a full-stack practice platform built around problems that look like real systems instead of interview puzzles.

Earlier work leaned more toward tooling and the web: Tetrapack, a parser that renders Notion-style blocks in isolation, and folder-routing, an Express wrapper that maps a directory tree to routes. Both came out of wanting a thing that did not exist yet in the shape I wanted it.

Models, from scratch

I learn architectures by implementing them rather than reading about them. That started with a DCGAN built from the 2015 Radford et al. paper, and has since moved through hierarchical reasoning models, masked diffusion and flow matching — training runs small enough to fit in Colab and instructive enough to be worth the GPU hours.

The classical work lives on Kaggle too: EDA and regression on 80 Cereals and Palmer Penguins, convolutional classifiers for leaf disease and cloud types, and sentiment analysis for the NLP side. My full profile is @ashishk1331.

Python, and things I wrote in it

Python is where I go when the problem is data-shaped. A scraped database about Bleach, a set of image algorithms implemented by hand, and a Brainfuck compiler are all in there. One line from the Zen of Python has held up better than most engineering advice I have read since:

If the implementation is hard to explain, it's a bad idea.
If the implementation is easy to explain, it may be a good idea.

Mobile and design

React Native and Expo were my first professional stack, and I still reach for them: Altar and Commit both shipped with a real data layer, a small design system and proper auth rather than screens glued together.

I design my own products before building them, which mostly saves me from building the wrong one. The case studies are public: Commit and Altar.

Elsewhere

I write regularly — see the notebook — and moderate JavaScript and React tags on Dev.to. I ran the Android track at GDG on Campus KNIT, and I have put together a couple of puzzle sets for people who like that sort of thing — Crack the Nut and 100 Days of Code.

Away from the keyboard it is mostly albums on repeat — Kanye, Kendrick, Tyler, and a long stretch of Billie's Hit Me Hard and Soft, which Finneas produced well enough that I wrote a whole post about his own record. I also like graphs, type systems, and arguing about video games.