For investors
Independent research lab.
Peer-reviewable output.
Skelf Research is an independent UK AI research lab with 19 current public products, 18 with a detected OSS licence. Five research pillars. One methodology: every hypothesis is a runnable, testable, peer-reviewable repository. The most-replicable, hardest-to-fake moat in open-source AI research.
What makes Skelf different
Independent research lab
The current public output is peer-reviewable software rather than a hosted commercial service. Engagements are evaluated against the lab's published research areas.
19 public products
18 currently carry a detected OSS licence. Maturity differs by project, and the portfolio is available for direct technical review.
Five research pillars
LLM Cognition & Prompt Theory, Safe & Verifiable Computing, Formal Optimisation & Decision Science, Edge Intelligence & On-Device AI, and Robotics & Autonomous Systems. Each pillar has multiple projects, an article series, and a community of contributors.
Methodology is the moat
Hypotheses as software. Every research question is encoded in a runnable, testable, falsifiable repository. This is the part that compounds — every year we have more artefacts that are public, peer-reviewable, and impossible to replicate without doing the work.
Geographic and assurance positioning
Skelf is a UK company. Several projects explore security, privacy, and auditability, but project features do not constitute organisational certification or regulatory compliance. Assurance is evaluated per engagement.
Multi-language stack
Rust, Go, Zig, Python, Dart, Lua, JavaScript, TypeScript. The thesis is language-agnostic — we choose the language for falsifiability, not fashion. The system is not a single point of failure.
Geographic focus
We target the markets where open-source AI infrastructure has the highest leverage: US, UK, EU, Canada, AUSNZ, Japan, China. Our 5,380 quarterly Google impressions (3 months) come predominantly from these regions; we convert almost none of them. That gap is the opportunity.
| Region | Quarterly impressions | Clicks | Target? | Focus |
|---|---|---|---|---|
| US | 3,256 | 0 | Yes | AI infrastructure, on-device, observability |
| UK | 133 | 3 | Yes | EU AI Act, GDPR, fintech, sovereign AI |
| India | 143 | 1 | No | (deprioritised; non-target region) |
| Canada | 155 | 0 | Yes | Sovereign AI, MILA, Toronto Vector Institute |
| Germany | 105 | 0 | Yes | EU AI Act, Max Planck, formal methods |
| France | 81 | 0 | Yes | INRIA, formal methods, sovereign AI |
| Australia | 59 | 0 | Yes | ANZ AI infrastructure and robotics |
| Japan | 50 | 0 | Yes | Sakana, Preferred Networks, formal verification |
| Netherlands | 48 | 0 | Yes | EU AI Act, A11Y, edge AI |
The portfolio
19 current public products, organised by technical domain. Click through for source, licence, documentation, and scope.
LLM & Agents
promptel · blogus · mpl · route-switch · anouk · direktor
Search & Retrieval
embedcache · memista · polymathy · slorg
Systems & Runtime
zviz · numaperf · gpuemu
Optimisation & Decision
savanty · compere · waremax
Privacy & Trust
perishable · l0l1 · tessera
Roadmap
- 2026 H1 19-product public portfolio, 5 research pillars, long-form research articles, and a technical glossary
- 2026 H2 3-5 new repositories, expanded article series on prompt theory and on-device AI, comparison articles for every project, /research taxonomy page
- 2027 H1 First external research collaborations; first papers co-authored with academic partners; expanded robotics pillar
- 2027 H2 Foundation / research-grant partnerships; second wave of projects in compliance + robotics
Risks and constraints
A brief that lists only strengths is not a brief. These are the constraints a reviewer would find anyway, stated first.
- Key person The lab is fewer than five people and the founder writes most of the code and all of the articles. Breadth across five pillars is a deliberate research choice and a concentration risk at this size. Mitigation is that every artefact is public and permissively licensed, so continuity does not depend on access to a private repository.
- No revenue engine Open research artefacts do not monetise themselves, and we have ruled out the two easiest routes — hosted services and closed cores — because both make the work less inspectable. Revenue comes from commissioned investigations with published prices (see services), sponsored research, and work packages in funded programmes. That model is new and unproven at our scale, and it is slower and lumpier than subscription revenue.
- Portfolio breadth Nineteen products across five domains means nineteen surfaces that can drift out of date. Several are explicitly experimental. A reviewer should read the per-project status and licence rather than treating the count as uniform maturity.
- Adoption is unproven Repository star counts are low single digits. The impressions table above shows the same thing from the demand side: the audience finds us and does not convert. We report this rather than substituting a vanity metric, and closing that gap is the nearest-term objective.
- Model-layer dependency Work above the model layer is insulated from which model is underneath, but not from the possibility that a frontier lab ships a capability that makes a component redundant. Small, narrowly-scoped artefacts limit the blast radius of that happening to one repository at a time.
Team
Dipankar Sarkar
Researcher · Founder
Independent AI researcher. Designs and implements the projects. Writes the articles. Edits the YAMLs.
Contact
- Research collaborations admin@skelfresearch.com
- Partnerships and suppliers admin@skelfresearch.com
- Press and media contact@skelfresearch.com
- GitHub organisation github.com/Skelf-Research