Custom development
Websites, web apps, internal tools, integrations and automations. Built to spec and shipped properly, not left as a prototype.
// dev collective
Webuildsoftwarethatships.
PulseOnix is a small collective of developers and researchers: client work, open source, AI engineering, hackathons.
Available for new projects
Four things, done properly.
The range is deliberately small: build for clients, push the frontier in AI and systems, publish in the open, and show up when it counts. Every pillar is backed by real work.
AI & ML engineering
Model fine-tuning, inference engines, agent pipelines and Hugging Face ecosystem work: from a weekend experiment to a production line.
Open source & research
Systems programming, AI research and developer tooling published in the open, with honest results and reproducible reports.
Hackathons
Fast, focused builds under pressure. An idea becomes a working demo in a weekend, and we show up with something real.
Built in the open, still moving.
Active research and open-source projects, currently in development and running locally. Scroll: each project files the previous one away.
research
Kestrel
A consequence-driven agent that learns hidden rules purely from experience, no corpus and no pretraining, across six phases from tabular world models to self-modifying exploration.
- Verified knowledge transfer between agents: 0% to 100% in one episode
- Specialists to generalist: 3.1x speedup, all rules mastered
- Self-modifying exploration: 30% faster re-mastery after rule swaps
- Runs on CPU in seconds, every claim backed by logged experience
in development
PRISM
A CPU LLM inference engine in C99 where everything is a plugin: kernels, model architectures, tokenizers, samplers, even the generation loop and frontend.
- The whole engine composes from a ~500 line registry core
- GGUF and llama2.c checkpoints, F32/F16/Q8_0/Q4_0 kernels
- AVX2 kernels parity-tested against the reference backend (about 9x f32)
- OpenAI-compatible REST API and quantized KV cache included
open source
hfm
A multi-account Hugging Face manager: one CLI and one MCP server for every HF account, repo and Space, built for humans and AI agents alike.
- OAuth device flow: no tokens pasted into dotfiles, auto-refreshed
- MCP server exposes the same operations to AI agents over stdio
- Read-only mode and an audit log as safety rails for agent access
- Repos, uploads, secrets and transfers, all account-switched
open source in development
vek
A systems-level project written in C, exploring low-level programming and performance-critical computation.
- Written in C for direct hardware control and memory management
- Systems programming focus: performance, efficiency, resource control
- Low-level implementation without runtime dependencies
Small crew, parallel skills.
A handful of friends who build together: systems and inference, AI and tooling, web and automation. We cover each other's gaps and hold everything we ship to the same bar.
Alish Karki
AI & Embedded Developer
AI agents, embedded systems & developer tools
@Alish-0x
Ritesh Raut
AI & Full-Stack Developer
AI agents, ML & web apps
@RautRitesh
Sewak Sunar
Mechanical Developer
Computational design & engineering tools
@sewaksunar
Sudin Rai
AI & Full-Stack Developer
AI agents, RAG systems & tooling
@sudin-techHave a project in mind?
Tell us what you're trying to build. We'll reply with honest feedback, a realistic scope, and a clear path, even if it's "you don't need us for this."
[email protected]Currently accepting new projects