Good questions. Traceable evidence. Human judgement.
A public view of TalkToAI's research direction: AI-assisted workflows that connect a question to its sources, make assumptions visible and keep people responsible for the result.
Research themes from the public archive, expressed as work to investigate and evaluate.
01 / HUMAN-CONTROLLED AI
Can a workflow show its work?
Explore how agents can help orient a project, inspect sources, plan a task and produce a useful output while keeping actions, assumptions and human review visible.
The practical interest is inspectable assistance, with explicit boundaries between a generated draft, an executed action and a verified result.
The earlier briefing explored literature discovery, source tracking, topic-expert identification and AI-assisted research writing. A useful output should preserve its evidence trail and make missing support visible.
Drafting can help structure a paper or technical brief. Citations, methods and conclusions still need a person's review.
Security, encryption concepts and quantum-readiness were part of the historical scoping context. The constructive next step is a defined question, a baseline and an evaluation plan.
This overview makes no claim of proven security, quantum advantage, hardware performance or production deployment.
A proposed six-step workflow. Each step should leave something concrete that another person can inspect.
Orient
Write the question, intended user, scope and exclusions. Define what would count as a useful answer.
Inspect
Collect public sources or authorised project material. Record where it came from and distinguish source statements from inference.
Plan
Choose a baseline, method, evaluation conditions and acceptance criteria before interpreting the outcome.
Act
Create a bounded draft, analysis or implementation. Preserve the inputs and the steps needed to understand the work.
Verify
Review citations, assumptions, calculations and checks. State incomplete coverage and unresolved failures alongside the result.
Report
Publish a clear brief with evidence, limitations and next questions. A draft or passing check only establishes what it actually measures.
Make collaboration concrete
The archive sought structured technical critique and proof-of-concept scoping. These are useful starting documents for that conversation.
One-page question
A challenge statement
Describe the problem, who experiences it, the current approach and one improvement you want to assess. Keep the first exchange non-confidential.
Methods before claims
A review brief
List the proposed method, relevant public sources, assumptions and strongest uncertainties. Ask reviewers which evidence would change their assessment.
A bounded demonstration
An evaluation plan
Identify the baseline, dataset or fixture, conditions, metrics and reporting format. Distinguish feasibility from demonstrated performance.
Historical context. The earlier page discussed potential UK academic, compute and innovation routes, including UKRI, STFC, Hartree Centre, NQCC, Innovate UK and University of Nottingham research influence. These mentions are background from that archive; they do not establish a current relationship, eligibility or a recommendation for any programme.
Public work you can inspect
Start with the maintained source and documentation. The static website requires no research account or confidential upload.
Research repository
Read public papers and working material in their source context. Evaluate each document's methods, status and stated limits.
For a research discussion, share the public problem statement, the material you can disclose and the kind of technical review you need. A focused question makes the next step easier to scope.