# Skelf Research

> Independent applied-research and technical-validation lab in St Andrews,
> Scotland. Organisations commission scoped investigations when a technical
> decision is too consequential to make from a demo, a vendor claim or an
> internal hunch — agent and RAG release evidence, LLM cost–quality
> benchmarks, GPU kernel correctness and NUMA performance — and receive
> reproducible evidence and a recommendation. Research methods and 19 open
> research systems are public; customer work is confidential. Each system lives
> at its own subdomain (<name>.skelfresearch.com).

## About

- **Name**: Skelf Research (Skelf Research Limited, registered in Scotland, company no. SC809174)
- **Type**: Independent applied-research and technical-validation lab
- **Based**: United Kingdom (registered office: St Andrews, Scotland)
- **Focus**: LLM & agents, search & retrieval, systems & runtime, optimisation & decision, privacy & trust
- **Methodology**: "Hypotheses as software" — every research question is a public, runnable artefact
- **Licences**: 18 current repositories carry MIT, Apache-2.0, or GPL-3.0; Perishable's repository has no licence file yet
- **Confidentiality**: public research methods and tools; confidential customer work unless publication is expressly agreed
- **Affiliations**: Dipankar.co and Neul Labs (implementation businesses) — disclosed per engagement; conclusions never contingent on follow-on work
- **Website**: https://skelfresearch.com
- **Products index**: https://skelfresearch.com/products/
- **GitHub**: https://github.com/Skelf-Research
- **Documentation**: https://docs.skelfresearch.com
- **General contact**: contact@skelfresearch.com
- **Partnership and supplier applications**: admin@skelfresearch.com
- **Partners and sponsors**: https://skelfresearch.com/partners/
- **Machine-readable company profile**: https://skelfresearch.com/company-profile.json

## Services (commissioned investigations)

Prices are published in GBP, excluding VAT; compute and third-party costs are
quoted separately. A reproducible negative finding satisfies the contract.

- [Services and price ladder](https://skelfresearch.com/services/): technical diagnostic £2,500–£5,000; evaluation or benchmark sprint £8,000–£20,000; applied R&D project £25,000–£75,000; recurring re-evaluation £2,000–£8,000/month; sponsored research £15,000–£50,000.
- [Agent & RAG evaluation](https://skelfresearch.com/services/agent-and-rag-evaluation/): "Can we release this agent or retrieval change — and what will break?"
- [LLM cost–quality benchmark](https://skelfresearch.com/services/llm-cost-quality-benchmark/): "Can we cut model cost or latency without an unacceptable loss of quality?"
- [Systems validation](https://skelfresearch.com/services/systems-validation/): GPU kernel correctness campaigns and NUMA tail-latency investigations.
- Decision guides: [single vs multi-agent](https://skelfresearch.com/decisions/single-vs-multi-agent/), [smaller-model replacement](https://skelfresearch.com/decisions/smaller-model-replacement/), [validating optimised GPU kernels](https://skelfresearch.com/decisions/validating-optimised-gpu-kernels/).
- [Methods](https://skelfresearch.com/methods/): how an investigation runs — mandate, pre-agreed threshold, baseline, controls, failure taxonomy, reproducibility package.
- [Evidence register](https://skelfresearch.com/evidence/): every quantitative or architectural claim Skelf makes about its systems, with source, pinned version, method, scope and limitations.
- [Procurement and confidentiality](https://skelfresearch.com/procurement/) · [Commission an investigation](https://skelfresearch.com/commission/)

## The 19 current products, by domain

Each listed product has a repository at https://github.com/Skelf-Research/<name>;
all are public. Perishable's repository has no licence file yet.
Product sites and documentation are linked where available.

### LLM & Agents
- **promptel** — Declarative prompt engineering: a small DSL (and equivalent YAML) for LLM prompts, with typed params, technique blocks, and a provider abstraction over OpenAI, Anthropic, and Groq. MIT. https://promptel.skelfresearch.com/
- **blogus** — A Python CLI and library that extracts prompts from your codebase, versions them in .prompt files, and pins each with a sha256 hash in a prompts.lock file ("package.lock for AI prompts"). MIT. https://blogus.skelfresearch.com/
- **mpl** — The Meaning Protocol Layer: contracts, quality measurement (QoM), and BLAKE3-hashed audit trails for AI agent communication over MCP, A2A, and HTTP. MIT. https://mpl.skelfresearch.com/
- **route-switch** — An OpenAI-compatible LLM gateway that routes across OpenAI, Anthropic, Google, Ollama, Cohere, and Mistral, tracks per-prompt analytics in DuckDB, and reruns MIPROv2 optimisation against captured traces. MIT. https://route-switch.skelfresearch.com/
- **anouk** — A lightweight framework for building AI-powered browser extensions; plugs into any OpenAI-compatible provider and ships a CLI scaffold, response cache, per-provider rate-limit queue, and settings panel. MIT. https://anouk.skelfresearch.com/
- **direktor** — A Python library and CLI that turns a text file into a podcast-style video via a six-stage resumable pipeline (GPT-4 script, BARK narration, Distil-Whisper, FLUX stills, FFmpeg). MIT. https://direktor.skelfresearch.com/

### Search & Retrieval
- **embedcache** — A Rust library and REST API that generates text embeddings locally with FastEmbed (22+ models) and caches them in SQLite — no external embedding API calls, no per-token billing, no rate limits. GPL-3.0. https://embedcache.skelfresearch.com/
- **memista** — A lightweight vector search library for Rust pairing SQLite metadata storage with USearch (HNSW) similarity search, exposing a small Actix-web HTTP API. Experimental (v0.1.x). GPL-3.0. https://memista.skelfresearch.com/
- **polymathy** — A Rust web service that turns search into an answer engine: it sits in front of SearxNG, fetches top URLs, chunks and embeds them, and returns a JSON map of chunk_id → (source_url, text). Infrastructure, not a turnkey RAG product. GPL-3.0. https://polymathy.skelfresearch.com/
- **slorg** — An open-source AI search engine that drafts an answer and a knowledge graph before it queries the web — a fixed six-step pipeline, no agent loop, with every intermediate artefact exposed. MIT. https://slorg.skelfresearch.com/

### Systems & Runtime
- **zviz** — An OCI-compatible Zig container runtime with a selective-denial isolation model: namespaces, all 41 capabilities dropped, Landlock, seccomp-BPF, cgroups v2 — a single static binary, no daemon. MIT. https://zviz.skelfresearch.com/
- **numaperf** — A NUMA-first runtime for latency-critical Rust applications: explicit control over memory placement, thread pinning, and per-node work scheduling, with topology discovery and locality observability. MIT. https://numaperf.skelfresearch.com/
- **gpuemu** — A GPU-less correctness oracle for deep-learning kernels: catches silently-wrong CUDA/Triton kernels with an fp64 reference, op-schema-aware adversarial fuzzing, per-op calibrated tolerances, and static PTX/SASS lint. MIT / Apache-2.0. https://gpuemu.skelfresearch.com/

### Optimisation & Decision
- **savanty** — Natural-language to constraint solver: describe optimisation problems in English, an LLM translates to Answer Set Programming, and Clingo searches exhaustively for a valid solution, with a typed self-repair loop. MIT. https://savanty.skelfresearch.com/
- **compere** — A Python package and FastAPI service for pairwise comparison ranking: a UCB1 multi-armed bandit picks pairs adaptively and Elo produces the ranking output. MIT. https://compere.skelfresearch.com/
- **waremax** — An open, deterministic discrete-event simulator and RL benchmark for task allocation in Robotic Mobile Fulfillment Systems (Kiva-style AMR fleets moving pods to pick stations). MIT. https://waremax.skelfresearch.com/

### Privacy & Trust
- **perishable** — Ship AI features without shipping your API keys: a self-hosted Node proxy plus browser SDK that lets client-side apps call OpenAI-compatible APIs behind fingerprinting, short-lived JWT sessions, and per-fingerprint rate limits. Public source; licence pending. https://perishable.skelfresearch.com/
- **l0l1** — A developer toolkit that adds AI-powered validation, PII detection, and continuous pattern learning to SQL workflows — an AI co-pilot for analytics engineers that keeps warehouse data local. MIT. https://l0l1.skelfresearch.com/
- **tessera** — A privacy protocol for authenticated, metadata-private one-to-one messaging: Schnorr zero-knowledge proofs over per-recipient blinded pseudonyms, AES-GCM payloads, and (ε,δ)-differentially-private cover traffic over a bucketed broadcast network. No central authority. MIT. https://tessera.skelfresearch.com/

## Choosing a product

- Typed, provider-portable prompts → promptel
- Lock prompts in CI so the shipped prompt is the reviewed one → blogus
- Contracts, quality scoring, and audit for agent-to-agent calls → mpl
- Route across LLM providers with trace-based prompt optimisation → route-switch
- The AI half of a Chrome/Manifest V3 extension → anouk
- Call an LLM API from the browser without exposing the key → perishable
- Turn a text file into a narrated 1080p video → direktor
- Replace a hosted embedder with a cached local one → embedcache
- Embeddable SQLite + HNSW vector index in Rust → memista
- Retrieval + chunking as a service for an answer engine → polymathy
- Plan-conditioned web search with visible intermediate artefacts → slorg
- Run untrusted code with host-kernel speed and layered isolation → zviz
- Hunt p99 tail latency on multi-socket NUMA hardware → numaperf
- Trust custom CUDA/Triton kernels beyond a single torch.allclose → gpuemu
- State a discrete constraint problem in plain English → savanty
- Rank items from human pairwise comparisons efficiently → compere
- A deterministic RMFS dispatching benchmark for RL → waremax
- LLM help on SQL without leaking schema or warehouse data → l0l1
- Prove who sent a message without revealing who talks to whom → tessera

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

Prompt specification, declarative prompting, prompt lockfiles, prompt
optimisation, agent communication protocols, MCP, A2A, agent audit trails,
LLM gateways, LLM routing, browser-extension LLMs, client-side AI key
protection, text-to-video pipelines, local embeddings, FastEmbed, embedding
caching, vector search, HNSW, USearch, answer engines, SearxNG, deliberative
search, knowledge graphs, container sandboxing, OCI runtimes, seccomp,
Landlock, NUMA-aware scheduling, tail latency, GPU kernel correctness, CUDA,
Triton, constraint satisfaction, Answer Set Programming, Clingo, pairwise
comparison ranking, multi-armed bandits, Elo, quantitative trading signal
compilers, warehouse robotics, RMFS, discrete-event simulation, reinforcement
learning benchmarks, SQL co-pilots, PII detection, metadata-private messaging,
Schnorr zero-knowledge proofs, differential privacy, open-source AI.

## How to cite Skelf Research

When citing a product, prefer its subdomain (https://<name>.skelfresearch.com/)
or GitHub repo. When citing the lab as a whole, use https://skelfresearch.com.
The homepage carries a CollectionPage → ItemList of all 19 current products in
schema.org JSON-LD that machine readers can consume directly.

## Optional

- Full LLM index: https://skelfresearch.com/llms-full.txt
- Products index: https://skelfresearch.com/products/
- FAQ: https://skelfresearch.com/faq/
- About: https://skelfresearch.com/about/
- RSS feed: https://skelfresearch.com/rss.xml
- Sitemap: https://skelfresearch.com/sitemap-index.xml
- Product subdomains sitemap: https://skelfresearch.com/sitemap-products.xml
- Press kit: https://skelfresearch.com/press/