Explore the work behind specialist AI: how it learns from experience, works across applications and is tested. Start with practical use-case guides, then read the original whitepapers behind the work.
Learning · execution · evaluation
Practical guides to the work behind the research. Explore the opportunity, the workflow and what a useful pilot should prove.
Explore the blogA claims warning should begin a focused review. See how an AI advisor can connect the warning, the governing guideline and the evidence a person needs to make a decision.
Billing work depends on process knowledge as well as data entry. Explore a measured path from recorded expert work to an AI specialist that can be evaluated on the decisions the workflow requires.
Start with one repeatable process, capture the work with consent and test whether a specialist can handle cases it has not seen before.
Start with the research behind specialist AI and real workflows. Explore the wider collection across enterprise intelligence and medical AI.
23 papers
Ordered by relevance to Digital HumanA Staged SFT + GRPO Training Methodology for an Agentic Case-to-Invoice Foundation Model
A staged method for teaching a model a billing workflow from recorded expert work. Supervised fine-tuning establishes task knowledge; reinforcement learning introduces process-based feedback for refinement.
READ THE PAPER HERE A closer look at DesktopTrace, which turns screen activity into meaningful training records by identifying interface controls, reconciling typed actions, and limiting which applications contribute data.
The preparation pipeline that turns recorded work into usable training examples, connecting actions to goals and process knowledge while checking that evaluation cases and answers stay separate.
An architecture that separates planning, execution and control so agents can work across existing applications, with explicit goals, bounded actions and recorded handling of failures and exceptions.
A prototype that translates claims-system warnings into guideline-linked checklists, helping reviewers identify missing information and permissible options. Evaluation uses synthetic claims, with the final decision remaining with a person.
A staged approach to reading insurance statements and provider contracts: a layout model locates relevant regions, then a language model converts those regions into structured, auditable records.
How circuit-design traces train a specialist model, with simulator feedback guiding improvement and a larger model joining when progress stalls. The paper examines the hybrid system’s engineering results.
How complete circuit-optimization episodes are collected and checked against a simulator, creating traceable training examples that preserve proposals, measurements, corrections and the evidence behind each design target.
Proposes a governed learning loop that turns user corrections and usage signals into improvements across knowledge, search, instructions, memory and model training, with evaluation and human approval before release.
Explores lightweight training of an enterprise search model using synthetic examples, testing whether adapting retrieval to a company’s vocabulary improves the passages found and the grounding of generated answers.
A memory architecture that separates personal preferences from organisational policy, resolves conflicting information by type and scope, and links authoritative knowledge to current source documents rather than storing stale copies.
Describes how Corporate GPT combines specialist agents, document search, source citations and permissions to answer enterprise questions, with human document selection, traceability and limits on retrieval.
Explains how document permissions and metadata travel from source systems into retrieval, so answers, citations and connector searches respect user access and changes to permissions can be traced.
How Corporate GPT prepares spreadsheets, slides, scanned pages and other complex files for search, preserving document structure and building summaries so important context survives the journey into an enterprise assistant.
A framework for delegating platform administration to AI agents, covering identity, knowledge, connectors and audit trails, with human approval and strict access controls around consequential changes.
An architecture for multilingual voice assistants that coordinate spoken answers with on-screen information, using native audio models, isolated sessions and a compact knowledge catalog to support responsive public-kiosk interactions.
Shows how structured schemas, staged extraction and batch processing turn radiology reports into consistent labels for training and evaluation. The paper reports an internal study of accuracy, reliability and processing cost.
Explores how a shared mammography model learns from large image collections, adapts to different findings, and is evaluated for transferability. The paper positions the model as a developer tool requiring further validation.
Describes how language generation, image localisation and explicit rules can be combined into traceable draft mammography findings. It examines disagreements between models, auditability and the evaluation still needed before clinical integration.
Explores a shared chest X-ray model that can be adapted to many findings. It documents training data, staged learning and controlled internal comparisons with other medical imaging models under the same evaluation conditions.
Examines a reusable three-dimensional model for chest CT studies, trained with paired images and radiology reports. It describes data preparation, efficient adaptation and internal validation across a panel of findings.
Describes a model that generates draft findings from chest X-ray images, including dataset curation, training and held-out evaluation. It is intended for downstream development and validation by medical device manufacturers.
Combines draft report generation with dedicated measurements for lung nodules and the aorta. The paper describes a research pipeline, its internal evaluation and specific gaps that require further work before clinical use.
Original papers, available to read here or download. Each paper describes its own methods, results and limitations.
See how the work was learned, how it was tested and what the result means.
Keep the situation, the decision, any correction and the result. Together, they form a useful example to learn from.
Consented demonstrations · architectureContext · decision · outcome
Defined success · recorded errors
Evaluation design principle. The engineering study uses held-out circuits; each published evaluation states its own method.
Test on work kept separate from training. Define success in advance and report the errors as well as the result.
Evaluation design principle · held-out engineering circuitsContext · decision · outcome
Defined success · recorded errors
Evaluation design principle. The engineering study uses held-out circuits; each published evaluation states its own method.
Keep the method and limits with every comparison. Qualified teams can review the supporting evidence with our model team.
Internal evidence · independent review remains a next stepContext · decision · outcome
Defined success · recorded errors
Evaluation design principle. The engineering study uses held-out circuits; each published evaluation states its own method.
Context · decision · outcome
Defined success · recorded errors
Evaluation design principle. The engineering study uses held-out circuits; each published evaluation states its own method.
Consented demonstrations · architecture
Explore research in progress, the wider architecture and the filed method. Detailed diligence materials are available by request.
Evaluating documented procedures against observed execution, and identifying the gaps that make automation fail.
Expert-weighted synthesis, compliance-first rewards, verified outcomes, and recovery learned from real work.
The compact multimodal model, training inside the customer boundary, evaluation methods, and the role of abstention.
Consented capture and compliant expert demonstrations as the foundation for a specialist model. International filings are in progress.
Why learning from everyone can mean learning from no one.
Read the perspective RESEARCH NOTEWhat changes when enterprise AI leaves the demonstration.
Read the perspective RESEARCH NOTEWhy the most valuable knowledge is difficult to describe.
Read the perspective