Applied AI Lab

Working systems, not concept art.

Every build publishes the problem, baseline, stack, time, cost, result, failure modes and production caveats. If a project is still in development, this page says so plainly.

Build log format

How every build is documented

Problem

What repetitive or expensive task exists?

Prototype

What was built and with which models/tools?

Measurement

What changed versus baseline?

Limitations

Where can it fail and what requires human review?

Build 01

ShieldedID — zero-knowledge identity protocol

Open-source, zero-knowledge identity protocol. Prove who you are, or just that you're over 18 or KYC-verified, without handing over a passport, address or date of birth to a third party.

Problem

Digital identity is broken: organisations collect scans of passports and personal data they don't need, creating honeypot databases and breach risk.

Prototype

Zero-knowledge proofs (Bulletproofs) running in WASM on-device. Pairwise subject IDs so users can't be tracked across services. Apache 2.0, self-hostable, no blockchain required, no trusted setup.

Measurement

Project-reported: 365+ tests, 91% coverage (internal test suite). No PII stored by design. These are internal results, not independent certification.

Limitations

Adoption depends on verifier integration and trust frameworks. Not a substitute for legal identity assurance where that is required. Human review still needed for consequential decisions. No independent security audit or certification is claimed here.

Status: open source, in active development

v1.5.0, Apache 2.0. Test counts and coverage are project-reported (internal), not independently verified. Relevant to age verification (Online Safety Act) and the UK digital identity trust framework, but any production use should follow your own security review and due diligence. View on GitHub →

Build 02

STCL — Semantic Token Compression Layer

Intelligent, provider-agnostic semantic token compression. Reduces LLM input token usage by ~38.2% on average (up to 59.4%) while preserving meaning. Open source, in active development.

Problem

LLM API costs scale with input tokens. Most teams pay for redundant, low-information content in every prompt.

Prototype

A pre-inference semantic compression layer, not a tokenizer. Runs before the LLM, so no model changes or retraining. Works with OpenAI, Anthropic, Google, Mistral and Groq. TypeScript, Node 18+, Docker-ready.

Measurement

Project-reported: average input reduction 38.2%, best case 59.4%. Latency overhead ~3.5ms. 74/74 internal tests passing. Deterministic and auditable.

Limitations

Compression is lossy by design; quality depends on content type. Always validate compressed output for consequential use. Human review still required for high-stakes decisions. Figures are project-reported, not independently benchmarked.

Status: open source, in active development

MIT licensed. Performance figures are project-reported (internal), not independently benchmarked. View on GitHub →

Build 03

UnitedLines — encrypted social network

A censorship-resistant social network built for free expression. Signal-grade encryption, transparent moderation, no ads and no tracking. Installable as a PWA.

Problem

Mainstream platforms combine surveillance economics with opaque, silent moderation. Users get no privacy and no visibility into how their content is treated.

Prototype

An encrypted social network (PWA) with signal-grade end-to-end encrypted DMs, a client-side encrypted file vault, a public append-only moderation ledger, and reputation/credibility signals to counter sockpuppetry.

Measurement

Live at unitedlines.co.uk. No ads, no tracking, private by design. Some features are intentionally visible only to registered members.

Limitations

Encryption and censorship-resistance are technical properties, not guarantees of content quality. Moderation is transparent but still requires human judgement. Adoption depends on network effects.

Status: live

Live at unitedlines.co.uk. Source repository is private; access is by invitation.

Future projects

What's next in the lab

The lab prioritises small, useful tools that save a Scottish business real time over grand presentations with no deployable result.

AI Workflow Audit

Process map, ranked opportunities, risk/data check and an ROI test plan for one bounded workflow.

AI Implementation Sprint

One prototype workflow with SOP, human review and a measurement plan.

Team AI Briefing

A tailored practical session for a Scottish team, not a generic keynote.