Independent Research Lab · United Kingdom
We investigate the foundations of machine reasoning, computational intelligence, and safe AI systems — then publish everything. Our research artefacts span LLM cognition, formal optimisation, edge inference, and high-performance computing.
Research Areas
Each research area addresses an open problem at the frontier of AI and computational science.
How do we formalise the relationship between prompt structure and model behaviour? We study declarative prompt specification, automatic optimisation, and routing — treating prompts as first-class research objects rather than ad-hoc strings.
What does it take to run AI-generated code safely? Our research into memory-safe language design, container sandboxing, and NUMA-aware scheduling explores the systems foundations needed for trustworthy autonomous computation.
Can natural language interface with mathematical solvers? We investigate the bridge between human intent and formally provable solutions — from constraint satisfaction to quantitative signal compilation and intelligent ranking algorithms.
What are the limits of local inference? We study on-device LLM execution, mobile agent architectures, and privacy-preserving AI — exploring how much intelligence can live at the edge without any cloud dependency.
Research Artefacts
Every repository is both a research contribution and a usable tool — MIT or GPL-3.0 licensed for the community.
Investigating unified local LLM serving. A drop-in Ollama replacement exploring model management and inference unification.
↗Research into cost-quality optimisation for LLM routing. Implements MIPROv2-based automatic prompt tuning and model selection.
↗Exploring declarative prompt specification as a formal language. Write once, run anywhere — treating prompts as portable, typed artefacts.
↗Studying prompt lifecycle management. Extracts prompts from codebases and versions them as first-class dependencies.
↗Research into persistent agent memory architectures. Structured, queryable memory systems for long-running LLM agents.
↗Investigating ephemeral credential models for AI APIs. Scoped, time-limited token proxies for secure LLM access.
↗Exploring multi-modal generation pipelines. Text-to-video synthesis combining LLM scripting with generative media models.
↗Studying AI-assisted data analysis with formal validation. An SQL co-pilot that learns query patterns while preserving privacy.
Research into minimal-overhead sandboxing for untrusted code execution. Lightweight container isolation in pure Zig.
↗Investigating topology-aware scheduling for latency-critical workloads. NUMA-first memory allocation and thread placement in Rust.
↗Exploring memory-safe C dialects for AI code generation. A compiler research project targeting agent-written systems code.
↗Studying caching strategies for high-dimensional vector computations. Eliminates redundant embedding recomputation at scale.
↗Research into lightweight vector similarity search. SQLite-backed approximate nearest neighbour retrieval in pure Rust.
↗Investigating the transformation of keyword search into semantic answer generation. Async chunking and embedding pipelines in Rust.
↗Exploring programmable database architectures with embedded scripting. Lua-native data storage for AI workflow prototyping.
↗Liath reimplemented in Rust with RocksDB. Studying performance characteristics of pluggable storage engines with Lua query interfaces.
Research into compiling quantitative trading signals from visual specifications. From alpha hypothesis to verified executable in minutes.
↗Bridging natural language and constraint satisfaction. Describe optimisation problems in English, receive mathematically guaranteed solutions via formal solvers.
↗Studying efficient ranking under sparse feedback. Multi-armed bandit algorithms for achieving better orderings with fewer pairwise comparisons.
Investigating on-device LLM inference limits. Running full language models on mobile hardware via Flutter with zero cloud dependency.
↗Exploring mobile-first agent architectures. On-device AI agents that browse, observe, and automate tasks autonomously.
↗Studying LLM integration patterns for browser extensions. A framework for rapid development of AI-augmented web experiences.
↗Research into deliberative search interfaces. A SvelteKit engine that reasons about query intent before retrieving results.
Our Approach
Skelf Research operates at the boundary between academic inquiry and real-world systems. We believe the most important questions in AI today — about reasoning, safety, efficiency, and privacy — are best answered by building working prototypes and publishing everything.
Our methodology is simple: identify an open problem, construct a hypothesis as software, stress-test it against real workloads, and release the results. Every repository is a peer-reviewable experiment.
Hypotheses as Software
Each project encodes a research question. The codebase is the proof — runnable, testable, and falsifiable.
Open Science by Default
24 public repositories. Every experiment is reproducible, every finding is auditable by the global research community.
Systems-Level Rigour
We choose Rust, Zig, and Go not for fashion but for falsifiability — deterministic performance makes claims measurable.
Privacy as a Research Constraint
On-device inference and zero-trust architectures aren't add-ons — they're design constraints that shape better science.
Collaborate
We welcome academic collaborators, research partners, and funders who believe the hardest problems in AI deserve open, rigorous, reproducible investigation.