Solo founder and independent ML researcher based in Rajkot, Gujarat. I build systems infrastructure and do original research — simultaneously, not sequentially. I also work as a creative designer and brand strategist — I've been shaping how businesses look, feel, and position themselves in the market since I was 16.

I work across three domains at once: production AI systems, original ML research, and my own database infrastructure. The common thread is complexity — how it emerges, how it can be modelled, and how to build systems that don't collapse under it.

My research sits at the intersection of pangenomics and graph neural networks. My infrastructure work is about building the query and storage layer I actually want to use. Neither is a side project. Both are the job.

I'm affiliated with Darshan University as an independent ML researcher and completing my BTech in 2026. No IIT, no metro city, no institutional lab. That's a feature, not a bug — it means I chose what to build rather than inheriting a problem to solve.

The design work runs alongside everything else. I started doing market image and brand positioning at 16 — long before the systems work — and it never stopped. Visual identity, creative direction, how a product earns trust before someone reads a single line of copy. That eye for how things look and feel shapes how I build too.

Interests span more than one domain by design: pangenomics, computational neuroscience, astrophysics, quant finance. The framing is complexity science across different scales — molecular, cosmic, economic.

• CoffeeQL. A Rust-native universal query orchestration engine — one interface across PostgreSQL, MySQL, MongoDB, and Redis. Published to npm and PyPI via PyO3/maturin. v0.3.1 is out with 265/265 tests passing. v0.4.0 in progress. Open sourcing the lexer, parser, and error layer under Apache 2.0 + Commons Clause; the planner, executor, and adapters stay closed.

• Something else. CoffeeQL isn't being built in isolation. There's a larger system it's designed to feed into — something that's been architected quietly for a while now. Not ready to talk about it yet. The shape of it will make sense once it ships.

• Ariadne / PanGNN. A Path-Aware Graph Attention Network for pangenome read alignment. Trained on HPRC v1.0 chromosomes 1–19. F1 of 94.5% on held-out chr20–22 and 96.0% on the independent HGSVC3+HPRC 2024 Joint Graph from EBI — +10.8 percentage points over VG-Giraffe. Published to Zenodo (10.5281/zenodo.21167336). bioRxiv preprint in preparation.

ariadne Path-Aware Graph Attention Network for pangenomics. Trained from scratch, validated against two independent graphs, beats the leading alignment tool by over ten points. The model I built when the existing tools weren't good enough.

CoffeeQL (moving to its own org) Rust-native query orchestration across four databases under a single interface. The part I'll open source is the language layer — the rest earns its keep by powering BobaDB.

[ redacted ] (not yet) Every architectural decision in CoffeeQL points somewhere. That somewhere has a name, a design doc, and a timeline. It'll be public when it's ready to be public — not before.

Nothing here is a tribal preference. It's the set of tools I've debugged enough to know how they fail.

• Build the hard thing first. Infrastructure that only works under ideal conditions isn't infrastructure.
• Instrument before optimising. If I can't see it, I assume it's broken and I'm just not being told.
• Research and product in parallel, not in sequence. Waiting until the paper is done to start building the system costs a year.
• The unconventional path is only a disadvantage if you let other people define what conventional is worth.

• GitHub — KhushviB
• Linkedin — Khusshvii
• Research — Zenodo: 10.5281/zenodo.21167336
• Happy to talk about: pangenomics, graph ML, Rust systems design, database architecture, independent research, and building outside the usual institutions.

_{Opinions are mine. Code is MIT unless a repository says otherwise.}