# Skelf Research — Full LLM Index

> Independent AI research lab based in the United Kingdom. Skelf Research turns
> sharp research questions into runnable software, shipping a family of 18
> focused public products for AI builders. Each product does one thing
> well, is explicit about its scope boundaries, and lives at its own subdomain
> (<name>.skelfresearch.com). This document is the extended, machine-readable
> index; the short version is at https://skelfresearch.com/llms.txt.

## The thesis

Most AI tooling is either a sprawling framework that tries to do everything or
a hosted product that hides its internals. Skelf Research takes the opposite
stance: build small, honest tools, publish them in full, and let people compose
them. The methodology is "hypotheses as software" — every open question becomes
a runnable, testable, peer-reviewable repository rather than a paper on a shelf.
The code is the proof.

The portfolio clusters into five domains:

1. LLM & Agents — prompts, agent protocols, routing, and client-side LLM plumbing.
2. Search & Retrieval — local embeddings, vector search, answer-engine infrastructure, deliberative search.
3. Systems & Runtime — container sandboxing, NUMA-aware runtimes, and correctness oracles.
4. Optimisation & Decision — constraint solving, ranking, trading-signal compilation, and simulation.
5. Privacy & Trust — metadata-private messaging, API-key protection, and SQL data protection.

## Organisation facts

- Legal name: Skelf Research Limited
- Company number: SC809174 (registered in Scotland)
- Registered office: 19 Drumcarrow Road, St Andrews, KY16 8SE, United Kingdom
- Founder: Dipankar Sarkar
- Website: https://skelfresearch.com
- GitHub org: https://github.com/Skelf-Research
- Docs: https://docs.skelfresearch.com
- General contact: contact@skelfresearch.com
- Partnership and supplier applications: admin@skelfresearch.com
- Partners and suppliers: https://skelfresearch.com/partners/
- Machine-readable company profile: https://skelfresearch.com/company-profile.json
- Licensing: 17 current repositories carry MIT, Apache-2.0, or GPL-3.0; Perishable's licence is pending

## Product reference (accurate, sourced from each product's own llms.txt)

For each: name — subdomain — language — licence — description — key scope notes.

### promptel
- Subdomain: https://promptel.skelfresearch.com/  ·  Repo: https://github.com/skelf-research/promptel  ·  Docs: https://docs.skelfresearch.com/promptel/
- Language: JavaScript  ·  Licence: MIT
- 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.

### blogus
- Subdomain: https://blogus.skelfresearch.com/  ·  Repo: https://github.com/Skelf-Research/blogus  ·  Docs: https://docs.skelfresearch.com/blogus/
- Language: Python  ·  Licence: MIT  ·  PyPI: blogus
- Extracts prompts from your codebase, versions them in .prompt files, and pins each with a sha256 hash in prompts.lock ("package.lock for AI prompts").
- Scope: a build-time tool. It does NOT run as a service, sit between app and model, or collect production traces. It makes the shipped prompt the reviewed prompt.

### mpl (Meaning Protocol Layer)
- Subdomain: https://mpl.skelfresearch.com/  ·  Repo: https://github.com/Skelf-Research/mpl
- Language: Rust  ·  Licence: MIT
- A protocol layer between AI agents and their tools (MCP, A2A, HTTP). Validates payloads against versioned semantic-type contracts, scores them against quality profiles (QoM), and writes BLAKE3-hashed, provenance-attributed audit records.

### route-switch
- Subdomain: https://route-switch.skelfresearch.com/  ·  Repo: https://github.com/Skelf-Research/route-switch  ·  Docs: https://docs.skelfresearch.com/route-switch/
- Language: Go  ·  Licence: MIT
- An OpenAI-compatible LLM gateway that routes across OpenAI, Anthropic, Google, Ollama, Cohere, and Mistral (via gollm), tracks per-prompt analytics in DuckDB, and reruns MIPROv2 prompt optimisation against captured production traces.

### anouk
- Subdomain: https://anouk.skelfresearch.com/  ·  Repo: https://github.com/skelf-research/anouk  ·  Docs: https://docs.skelfresearch.com/anouk/
- Language: JavaScript  ·  Licence: MIT  ·  npm: anouk
- A framework for building AI-powered browser extensions. Plugs into any OpenAI-compatible provider (OpenAI, Anthropic, Together.xyz, Ollama, Hugging Face) and ships a CLI scaffold, per-request response cache, per-provider rate-limit queue, and runtime settings panel. Targets Chrome/Chromium Manifest V3.

### direktor
- Subdomain: https://direktor.skelfresearch.com/  ·  Repo: https://github.com/Skelf-Research/direktor  ·  Docs: https://docs.skelfresearch.com/direktor/
- Language: Python  ·  Licence: MIT  ·  PyPI: direktor
- Turns a text file into a podcast-style video via a six-stage resumable pipeline: script (GPT-4), narration (BARK via Replicate), transcription (Distil-Whisper), image prompts (GPT-4), stills (FLUX-schnell via Replicate), and FFmpeg composition into a 1920x1080 MP4. Needs Python 3.11+, FFmpeg, OpenAI + Replicate keys, and an S3-compatible bucket.

### embedcache
- Subdomain: https://embedcache.skelfresearch.com/  ·  Repo: crates.io/crates/embedcache  ·  Docs: https://docs.skelfresearch.com/embedcache/
- Language: Rust  ·  Licence: GPL-3.0
- A Rust library and REST API that generates text embeddings locally using FastEmbed (22+ models: BGE, MiniLM, Nomic, multilingual E5) and caches them in SQLite keyed by content hash + model id. No external embedding API calls, no per-token billing, no rate limits.
- Scope: replaces the hosted embedder rather than memoizing it (distinct from CacheBackedEmbeddings-style wrappers).

### memista
- Subdomain: https://memista.skelfresearch.com/  ·  Repo: https://github.com/Skelf-Research/memista  ·  crate: memista
- Language: Rust (1.56+)  ·  Licence: GPL-3.0  ·  Status: experimental (v0.1.x)
- A lightweight vector search library pairing SQLite metadata storage with USearch (HNSW, IP metric, F32) and an Actix-web HTTP API (default 127.0.0.1:8083; POST /v1/insert, POST /v1/search, DELETE /v1/drop).

### polymathy
- Subdomain: https://polymathy.skelfresearch.com/  ·  Repo: https://github.com/skelfresearch/polymathy  ·  crate: polymathy  ·  Docs: https://docs.skelfresearch.com/polymathy/
- Language: Rust  ·  Licence: GPL-3.0
- A web service that transforms search into an answer engine: sits in front of a SearxNG instance, fetches top result URLs, chunks and embeds them, and returns a JSON map of chunk_id → (source_url, text). Single endpoint: GET /v1/search?q={query}. Infrastructure for building answer engines, not a turnkey RAG product.

### slorg
- Subdomain: https://slorg.skelfresearch.com/  ·  Repo/npm: npm:lorg  ·  Docs: https://docs.skelfresearch.com/slorg/
- Language: JavaScript (Node)  ·  Licence: MIT
- An AI search engine that drafts an answer and a knowledge graph before it queries the web. Fixed six-step pipeline (draft → extract graph → derive keywords → SearxNG query → fetch+extract → score 0–1); no agent loop. Exposes draft, graph, keywords, scored results, and token count.

### zviz
- Subdomain: https://zviz.skelfresearch.com/  ·  Repo: https://github.com/Skelf-Research/zviz
- Language: Zig  ·  Licence: MIT
- An OCI-compatible container runtime with a selective-denial isolation model: user/pid/mount/ipc/uts namespaces, all 41 capabilities dropped, Landlock LSM, seccomp-BPF, cgroups v2. 132 syscalls reach the host kernel at native speed, 24 denied at seccomp, socket() argument-filtered. Single static binary, no daemon.
- Scope: NOT a userspace kernel (unlike gVisor) and NOT a MicroVM (unlike Firecracker).

### numaperf
- Subdomain: https://numaperf.skelfresearch.com/  ·  Repo: https://github.com/Skelf-Research/numaperf  ·  Docs: https://docs.skelfresearch.com/numaperf/
- Language: Rust (MSRV 1.70)  ·  Licence: MIT
- A NUMA-first runtime for latency-critical Rust applications: topology discovery, RAII thread pinning, explicit memory placement (Bind, Preferred, Interleave, Local), per-node work scheduling, per-node sharded data structures, device locality, and observability. Linux x86_64/aarch64; graceful degradation on macOS.

### gpuemu
- Subdomain: https://gpuemu.skelfresearch.com/  ·  Licence: MIT / Apache-2.0
- Language: Rust CLI + Python client
- A GPU-less correctness oracle for deep-learning kernels. Catches silently-wrong CUDA/Triton kernels using an fp64 reference oracle, op-schema-aware adversarial fuzzing, per-op calibrated tolerances, and static PTX/SASS lint. Validation runs without a GPU; failures replay byte-for-byte from a seed. Install: cargo install gpuemu, pip install gpuemu.

### savanty
- Subdomain: https://savanty.skelfresearch.com/  ·  Repo: https://github.com/Skelf-Research/savanty
- Language: Python  ·  Licence: MIT
- Natural-language to constraint solver. An LLM (DSPy-orchestrated, default GPT-4o) translates English to a canonical assign(Var,Value) Answer Set Programming encoding; Clingo searches exhaustively over finite domains. Typed self-repair loop handles syntax_error / unsat / empty, including a minimal-unsat-core localisation.
- Fit: discrete constraint satisfaction (scheduling, assignment, timetabling, colouring). Not: continuous optimisation, ML, statistics, simulation.

### compere
- Subdomain: https://compere.skelfresearch.com/  ·  Repo: https://github.com/skelf-research/compere  ·  Docs: https://docs.skelfresearch.com/compere/
- Language: Python 3.11+ (FastAPI, SQLAlchemy)  ·  Licence: MIT  ·  pip: compere
- Pairwise comparison ranking. A UCB1 multi-armed bandit selects pairs adaptively; Elo produces the ranking. NOT Bradley-Terry, Thurstone, or TrueSkill.

### waremax
- Subdomain: https://waremax.skelfresearch.com/  ·  Repo: https://github.com/Skelf-Research/waremax
- Language: Rust core  ·  Licence: MIT
- 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). Scenarios are YAML. Determinism is tested: same seed + action sequence ⇒ byte-identical trajectory (ChaCha8 RNG, canonical tie-breaking).
- Scope: models ONLY pod-to-person RMFS dispatching; not AS/RS cranes, conveyor sortation, tugger trains, or human pickers.

### perishable
- Subdomain: https://perishable.skelfresearch.com/  ·  Repo: https://github.com/Skelf-Research/perishable  ·  npm: perishable
- Language: TypeScript/Node  ·  Licence: pending (public source; not currently described as open source)
- 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 without exposing the upstream key: fingerprinting, entropy collection, short-lived JWT sessions, per-fingerprint rate limiting. Works with OpenAI, Anthropic, OpenRouter, Ollama.
- Scope: NOT a hosted service, observability layer, per-user auth, or bot-proof — it raises the cost of casual abuse.

### l0l1
- Subdomain: https://l0l1.skelfresearch.com/  ·  Repo: https://github.com/Skelf-Research/l0l1
- Language: Python  ·  Licence: MIT
- A developer toolkit that adds AI-powered validation, PII detection, and continuous pattern learning to SQL workflows. Sits between an LLM and a warehouse: validates LLM-generated SQL against live schema, scans queries/prompts for PII before they leave the local process, and learns approved patterns per workspace. Connects to PostgreSQL, MySQL, SQLite, DuckDB; ships CLI, REST API, Jupyter magic, and VS Code/LSP.

## Navigation

- Homepage (aggregator hub): https://skelfresearch.com/
- Products index (18 current products, grouped): https://skelfresearch.com/products/
- About: https://skelfresearch.com/about/
- FAQ: https://skelfresearch.com/faq/
- Research approach: https://skelfresearch.com/research/
- Blog: https://skelfresearch.com/blog/
- Contact: https://skelfresearch.com/contact/
- RSS: https://skelfresearch.com/rss.xml
- Sitemaps: https://skelfresearch.com/sitemap-index.xml and https://skelfresearch.com/sitemap-products.xml