Independent UK AI research lab — we publish our research as software you can actually run.
Website · Docs · Research · Hugging Face
We don't write papers and leave the code as an exercise for the reader. Every project below is a runnable research artefact — install it, benchmark it, read the source, break it. Systems programming, LLM tooling, retrieval, agent protocols, privacy. All MIT / Apache-2.0.
gpuemu ·
RustYour CUDA/Triton kernel passestorch.allcloseand still ships a silent bug. gpuemu catches the ones that slip through — 10/10 LLM-style kernel bugs, 0 false positives, and no GPU required. site · docs · code
tessera ·
PythonMetadata-private messaging: hide who talks to whom, not just the payload. Schnorr zero-knowledge proofs + (ε,δ)-differentially-private cover traffic, no central authority. site · docs · code
sigc ·
RustThe Quant's Compiler. Take an alpha idea to a deterministic backtest and ship the same binary to production — in minutes, not sprints. site · docs · code
zviz ·
ZigRun code you can't trust, at native speed. 132 syscalls run native, 24 are blocked at seccomp, one is argument-filtered. No userspace kernel, no daemon. site · docs · code
mpl ·
RustA typed, measurable, provable contract layer between your agents and their tools — with tamper-evident audit trails that map to SOX, GDPR, and the EU AI Act. site · docs · code
slorg ·
PythonA search engine that drafts an answer and a knowledge graph before it queries the web. Open-source, OpenAI-compatible, SearxNG-backed. site · docs · code
🔍 Retrieval & memory
| Project | Links | |
|---|---|---|
| embedcache | Generate embeddings locally and cache them — no API keys, no rate limits | site · docs · code |
| memista | A lightweight vector search library for Rust | site · docs · code |
| polymathy | Turn search results into answer-engine chunks | site · docs · code |
🧩 LLM tooling & prompts
| Project | Links | |
|---|---|---|
| promptel | A declarative DSL for LLM prompts — write once, run across providers | site · docs · code |
| blogus | package.lock for AI prompts — version prompts like dependencies |
site · docs · code |
| route-switch | OpenAI-compatible LLM gateway with self-improving prompts | site · docs · code |
| perishable | Ship AI features without shipping your API keys | site · docs · code |
| anouk | Build AI-powered browser extensions in minutes | site · docs · code |
| direktor | AI text-to-video pipeline — turn articles into narrated videos | site · docs · code |
🤝 Agents & decisions
| Project | Links | |
|---|---|---|
| savanty | Describe an optimisation problem in English, get a guaranteed solution | site · docs · code |
| compere | Intelligent pairwise comparisons — better rankings with fewer votes | site · docs · code |
| l0l1 | An AI co-pilot for SQL — validation, PII detection, pattern learning | site · docs · code |
⚙️ Systems & performance
| Project | Links | |
|---|---|---|
| numaperf | A NUMA-first runtime for latency-critical Rust | site · docs · code |
| waremax | Deterministic simulator & RL benchmark for warehouse robotics | site · docs · code |
- Reproducible or it didn't happen. Every claim ships with the code and benchmark that back it.
- Runnable research. Each repo is both a paper's worth of ideas and a tool you can
installtoday. - Permissive by default. MIT / Apache-2.0, so you can build on any of it without asking.
Star what's useful, open an issue, or send a PR — every repo welcomes contributions. Building something with our work, collaborating on research, or funding it? contact@skelfresearch.com