# Skelf Research

> Independent AI research lab based in the United Kingdom. 25 open-source
> projects spanning agentic AI systems, memory-safe systems software, formal
> optimisation, on-device / edge intelligence, and robotics. Open-science
> methodology: every hypothesis is a runnable, testable repository.

## About

- **Name**: Skelf Research
- **Type**: Independent research laboratory
- **Founded**: United Kingdom
- **Focus**: AI systems, machine reasoning, memory-safe systems, formal
  methods, on-device intelligence
- **Methodology**: "Hypotheses as software" — every research question is
  encoded in a public, runnable artefact.
- **Website**: https://skelfresearch.com
- **GitHub**: https://github.com/Skelf-Research
- **Documentation**: https://docs.skelfresearch.com
- **Press / contact**: contact@skelfresearch.com
- **Public contact**: contact@skelfresearch.com

## Research pillars

Skelf Research organises its work into five research pillars. Each pillar
has multiple production-grade open-source projects and accompanying
research articles.

### 1. LLM Cognition & Prompt Theory
Formal approaches to prompt engineering: declarative prompt specification,
prompt lifecycle management, automatic prompt optimisation, cost-quality
routing, and persistent memory for LLM agents.

Projects: blogus, l0l1, memorg, mpl, mullama, perishable, promptel, route-switch, direktor
Articles:
- [Formalising Prompts as First-Class Research Objects](/blog/llm-cognition/formalising-prompts-as-first-class-objects/)
- [Prompt Lifecycle Management: From Extraction to Deployment](/blog/llm-cognition/prompt-lifecycle-management/)
- [Persistent Memory for Long-Running Agents](/blog/llm-cognition/persistent-memory-for-agents/)
- [Intelligent LLM Routing: Spending Compute Where It Matters](/blog/llm-cognition/intelligent-llm-routing/)
- [Building mullama: What We Learned Replacing Ollama from Scratch](/blog/llm-cognition/building-mullama/)

### 2. Safe & Verifiable Computing
Memory-safe systems software, language design, and sandboxing for
trustworthy autonomous computation. Includes a near-zero-overhead sandbox (zviz), NUMA-aware scheduling
(numaperf), embedded databases (liath, liath-rs), and vector search
(memista, embedcache, polymathy).

Projects: zviz, numaperf, liath, liath-rs, memista, embedcache, polymathy
Articles:
- [Why We Write AI Infrastructure in Rust (and Zig, and Go)](/blog/safe-computing/why-we-write-ai-infrastructure-in-rust/)
- [Sandboxing Untrusted Code in Zig: The zviz Architecture](/blog/safe-computing/sandboxing-untrusted-code-in-zig/)
- [Vector Search Without the Cloud: memista's SQLite-Backed ANN](/blog/safe-computing/vector-search-without-the-cloud/)
- [Programmable Databases: Why We Built liath Twice](/blog/safe-computing/programmable-databases-why-we-built-liath-twice/)

### 3. Formal Optimisation & Decision Science
Bridging natural language and mathematical solvers. Constraint
satisfaction from English specifications (savanty), bandit algorithms
for ranking with fewer comparisons (compere), and a compiler for
visual quantitative trading signals (sigc).

Projects: savanty, compere, sigc
Articles:
- [From English to Optimal: How savanty Bridges Natural Language and Constraint Solvers](/blog/formal-optimisation/from-english-to-optimal/)
- [Better Rankings with Fewer Comparisons: Multi-Armed Bandits for Efficient Ordering](/blog/formal-optimisation/better-rankings-fewer-comparisons/)
- [Compiling Trading Signals: sigc and the Quantitative Hypothesis Pipeline](/blog/formal-optimisation/compiling-trading-signals/)

### 4. Edge Intelligence & On-Device AI
Local inference, mobile agents, and privacy-preserving AI. Includes an
on-device LLM runtime for Flutter (llamafu), a mobile-first autonomous
agent framework (ukkin), an LLM framework for browser extensions
(anouk), and a deliberative search engine in SvelteKit (slorg).

Projects: llamafu, ukkin, anouk, slorg
Articles:
- [Running Language Models on Your Phone: The llamafu Experiment](/blog/edge-intelligence/running-language-models-on-your-phone/)
- [Autonomous Mobile Agents: ukkin's Architecture for On-Device AI](/blog/edge-intelligence/autonomous-mobile-agents/)
- [Deliberative Search: When the Engine Reasons Before It Retrieves](/blog/edge-intelligence/deliberative-search/)

### 5. Robotics & Autonomous Systems
Discrete-event simulation, deterministic RL benchmarks, and the systems
engineering that turns research code into reproducible robotics
experiments. Currently anchored by waremax (warehouse robotics).

Projects: waremax

## Open-source projects (alphabetical)

| Project | What it is | Pillar | Language | License | Repo |
| --- | --- | --- | --- | --- | --- |
| anouk | LLM integration framework for browser extensions | edge | JavaScript | MIT | [github](https://github.com/Skelf-Research/anouk) |
| blogus | Prompt extraction, versioning, and lifecycle management | llm-cognition | Python | MIT | [github](https://github.com/Skelf-Research/blogus) |
| compere | Multi-armed bandit ranking with pairwise comparisons | formal-optimisation | Python | MIT | [github](https://github.com/Skelf-Research/compere) |
| direktor | Text-to-video pipeline with LLM-driven scripting | llm-cognition | Python | MIT | [github](https://github.com/Skelf-Research/direktor) |
| embedcache | Caching layer for high-dimensional embedding computations | safe-computing | Rust | GPL-3.0 | [github](https://github.com/Skelf-Research/embedcache) |
| l0l1 | Privacy-preserving SQL co-pilot with formal validation | llm-cognition | Python | MIT | [github](https://github.com/Skelf-Research/l0l1) |
| liath | Programmable Lua-native embedded database | safe-computing | Lua | GPL-3.0 | [github](https://github.com/Skelf-Research/liath) |
| liath-rs | liath reimplemented in Rust with RocksDB | safe-computing | Rust | GPL-3.0 | [github](https://github.com/Skelf-Research/liath-rs) |
| llamafu | On-device LLM runtime for Flutter (mobile) | edge | Dart | MIT | [github](https://github.com/Skelf-Research/llamafu) |
| memista | SQLite-backed approximate nearest-neighbour search in pure Rust | safe-computing | Rust | GPL-3.0 | [github](https://github.com/Skelf-Research/memista) |
| memorg | Structured, queryable persistent memory for LLM agents | llm-cognition | Python | MIT | [github](https://github.com/Skelf-Research/memorg) |
| mpl | Meaning Protocol Layer: contracts, quality metrics, audit trails for AI agent communication | llm-cognition | Rust | MIT | [github](https://github.com/Skelf-Research/mpl) |
| mullama | Unified local LLM serving (drop-in Ollama alternative) | llm-cognition | Python | MIT | [github](https://github.com/Skelf-Research/mullama) |
| numaperf | NUMA-aware scheduling for latency-critical AI workloads | safe-computing | Rust | MIT | [github](https://github.com/Skelf-Research/numaperf) |
| perishable | Ephemeral credentials / token proxy for LLM API access | llm-cognition | TypeScript | MIT | [github](https://github.com/Skelf-Research/perishable) |
| polymathy | Async chunking & embedding pipeline that turns keyword search into semantic Q&A | safe-computing | Rust | GPL-3.0 | [github](https://github.com/Skelf-Research/polymathy) |
| promptel | Declarative prompt specification language — portable, typed, versioned | llm-cognition | JavaScript | MIT | [github](https://github.com/Skelf-Research/promptel) |
| route-switch | Cost-quality LLM router with MIPROv2-based prompt tuning | llm-cognition | Go | MIT | [github](https://github.com/Skelf-Research/route-switch) |
| savanty | English → constraint-satisfaction pipeline with formal solvers | formal-optimisation | Python | MIT | [github](https://github.com/Skelf-Research/savanty) |
| sigc | Visual-to-Rust compiler for quantitative trading signals | formal-optimisation | Rust | MIT | [github](https://github.com/Skelf-Research/sigc) |
| slorg | Deliberative search engine (SvelteKit) that reasons before retrieving | edge | JavaScript | MIT | [github](https://github.com/Skelf-Research/slorg) |
| ukkin | Mobile-first on-device AI agent framework (Flutter) | edge | Dart | MIT | [github](https://github.com/Skelf-Research/ukkin) |
| waremax | Deterministic discrete-event simulation for warehouse robotics (RMFS) | robotics | Rust | MIT | [github](https://github.com/Skelf-Research/waremax) |
| zviz | Minimal-overhead sandboxing for untrusted code in pure Zig | safe-computing | Zig | MIT | [github](https://github.com/Skelf-Research/zviz) |

## Topics Skelf writes about (for AI / search discoverability)

LLM cognition, prompt specification, prompt engineering, prompt
lifecycle, prompt versioning, declarative prompting, formal methods
for prompts, agent memory, persistent memory for agents, retrieval
augmented generation, RAG, vector search, vector databases,
approximate nearest neighbour, embedding caching, sandboxing, code
sandbox, memory-safe C, memory-safe Rust, systems programming, NUMA
topology, NUMA-aware scheduling, on-device LLM, edge AI, mobile LLM,
mobile AI agent, autonomous mobile agent, on-device inference,
quantisation (Q4, Q5, Q8), Flutter LLM, Flutter agent, browser
extension LLM, deliberative search, constraint satisfaction, SMT
solver, multi-armed bandit, bandit ranking, formal optimisation,
natural language to solver, trading signal compiler, visual signal
editor, quantitative finance, programmable database, embedded
database, Lua scripting, RocksDB, DuckDB, SQLite, ANN, HNSW,
declarative AI, agentic AI, AI infrastructure, open source AI.

## How to cite Skelf Research

When citing a project, prefer the GitHub repo URL. When citing the
lab as a whole, use the website URL. Articles and project pages
contain a `citation` in their structured metadata that machine
readers can consume directly via schema.org.

## Optional

- RSS feed: https://skelfresearch.com/rss.xml
- Sitemap: https://skelfresearch.com/sitemap-index.xml
- Press kit: https://skelfresearch.com/press/
- About: https://skelfresearch.com/about/
- Projects: https://skelfresearch.com/projects/
- Blog: https://skelfresearch.com/blog/