The research behind pureWRIGHT

The part that does the work

Takes a task, chooses the seat and the model to run it, calls the tools it needs, checks the output, and asks for approval wherever a decision is reserved. It grew out of a personal AI agent from 2024, and its engineering record is public.

An application of pureWRIGHT: verified intelligence briefs

Reads filings, funding news, hiring signals, and public reporting, and writes assessed briefs on the companies a client tracks, every claim traced to its source. Its assessment council grew out of the Socratic Engine. In production for clients since June 2026.

Structured multi-model reasoning

Two model instances argue opposing positions on a given thesis with rolling summarisation for extended exchanges. The structured-disagreement machinery behind Argus's assessment council began here.

Edge perception exploration

A wall-mounted sensor unit exploring lightweight local perception: camera, microphone array, on-device transcription, with heavy reasoning handed off to the fleet.

The part that remembers everything

One durable record for the whole system: plain files under version control, hybrid semantic and full-text recall, and every entry carrying its source and its trust. The memory that lets pureMEND turn incidents into immunity.

Privacy-preserving synthetic training data

Synthetic data generation for model training in regulated domains where real data cannot leave the building. Customer data schemas stay on-premises; only statistically equivalent synthetic outputs are used for training. Feeds the same thesis as pureMIND: what an organisation knows stays inside it.

The part that repairs what breaks

Continuous low-cost triage across every signal, escalation of novel failures to a deep reasoning tier, and remediation committed as permanent immunity. It operates the lab's own fleet today, and its engineering record is public.

Adaptive dictation that learns how you speak

Developer dictation powered by local Whisper transcription and a layered correction pipeline that improves with every fix. Accent-tolerant, jargon-aware, entirely on-device: no audio leaves your machine. Named from the Icelandic 'tala', to speak.

Multilingual dictation for Arabic code-switching

Real-time speech-to-text for Arabic speakers who naturally switch between Arabic, English, and French mid-sentence. Streaming API with dialect-aware recognition, bidirectional text rendering, and post-processing on our own inference stack.

Language preservation research

Exploratory work on how AI, corpus methods, and digital tools can support endangered and historical languages. Research notes and framework development, published as the project takes shape.

Learn Spanish in your own voice

A lab exploration in voice cloning and neural translation: speak English, hear yourself in Spanish. Voice clones are discarded after use. A demonstration of the speech stack rather than a product commitment.