35% reported hiring difficulty
DSIT’s 2025 survey reports 35% of surveyed organisations struggling to fill AI roles; 97% of respondents reported at least one skills gap. DSIT / G&T executive summary
Research note: operational delivery, methodological expertise, systems architecture, interdisciplinary synthesis and regulatory competence require different assessment methods. A role-specific requirement for production experience should not become a universal proxy for the full AI knowledge ecosystem.
Explore the source-grounded competence and recruitment analysis →
Understand the work, test the claims and choose your next move. Research engineering and safeguarding each need their own evidence.
A dedicated research-engineering workspace: experimental judgement, reproducibility and scaling evidence.
“critical bridge between theory and implementation”Google DeepMind — role definitions
DeepMind’s description connects research ideas with implementation and evaluation. Employer evidence
Suggested evidence: paper-to-code map, reproduced baseline, deviations from the paper, controlled comparisons and a report of unsuccessful runs.
DeepMind identifies model optimisation and distributed research infrastructure in research engineering. Employer evidence
Suggested evidence: training configuration, dataset provenance, seed sensitivity, compute budget, checkpoints and scaling bottlenecks. Fine-tuning methods should follow the actual role and research question, rather than a universal tool checklist.
Research engineers build systems to test and evaluate new ideas. Employer evidence
Suggested evidence: held-out tasks, ablations, uncertainty, leakage checks, failure slices and human adjudication. If using a model judge, validate agreement against human labels; a judge score alone is insufficient evidence.
The role includes implementation, optimisation and systems at scale. Employer evidence
Suggested evidence: measured latency and memory, reproducible environments, profiling, regression results and documented research-to-service limits.
DeepMind says research scientists normally hold a PhD; it does not impose that statement on every research engineer. Anthropic describes a mixed academic background and substantial overlap between engineering and research. A doctorate is valuable evidence of research training, not a universal job-title requirement. DeepMind · Anthropic hiring criteria
“engineers here do lots of research”
Suggested doctoral portfolio: original question → literature gap → defensible method → reproducible experiment → limitations → contribution. Explain negative findings, competing hypotheses and what would falsify your conclusion.
Suggested prompts, not an employer’s prescribed interview:
Architecture: Why this model, baseline and compute design? What alternatives did you reject, and with what evidence?
Signal path: Trace data collection → validation → representation → training → inference → decision. Identify leakage, uncertainty and human decision points.
Security: Explain trust boundaries, access, secrets, supply chain, adversarial inputs and containment. Show test evidence, residual risk and who can stop release.
Research contribution: Which improvement survives ablation, reruns and a fair comparator? What failed, and what did that teach you?
Agents are a system pattern, not a replacement for the human roles above. They sit around models, retrieval and tools, selecting steps and responding to results.
Suggested design: user goal → context and permissions → model-selected step → tool execution → verified observation → next step or stop. Keep an auditable trace and an accountable human owner.
RAG supplies retrieved context; a tool interface provides an action; neither automatically makes a system autonomous. The design should justify dynamic decisions rather than assume an agent is necessary.
Suggested division: research engineers investigate agent behaviour, evaluation and failure; ML / software engineers build reliable execution; architects define integration and authority; domain and safeguarding specialists assess risks to people. The organisation retains accountability for authorised actions.
Suggested tests: malicious retrieved instructions, unauthorised tool actions, stale context, fabricated success, repeated actions and runaway loops. Use synthetic data and sandboxed tools.
Suggested controls: least privilege, explicit action boundaries, approval for consequential steps, time and spending limits, verifiable tool results, stopping conditions, recovery and human escalation.
Anthropic — sandbox testing and stopping conditions · OWASP — application controls
Suggested extension: a literature-mapping agent could propose searches, retrieve approved documents, extract attributed claims and draft a synthesis. A human reviews sources, disagreements and the final contribution.
The existing browser demonstrations are learning tools; this guide does not claim they already run autonomous research agents. Measure task success, source accuracy, unsafe actions, cost and recovery against a simpler workflow before adding autonomy.
Employer descriptions are examples, not a universal hierarchy. Coding is a capability shared across several roles. The 15 role families below have overlapping skills and distinct responsibilities; careers links show employers to investigate, not proof of a current vacancy.
| Role | Primary emphasis in these examples | Evidence | Alignment with the supplied description |
|---|---|---|---|
| Research engineer | Implement, evaluate and scale experiments; connect theory and systems. | DeepMind | Strong alignment with research translation. Production ownership varies by team. |
| Research scientist | Develop hypotheses and new research directions; conduct experiments and publish. | DeepMind | “Purely theoretical” is too narrow; experimental work also matters. |
| ML engineer | Research-adjacent ML systems and tooling in G-Research’s example. | G-Research role | “Only uses existing APIs” is too restrictive. Examine each employer’s scope. |
| Software engineer | Reliable, scalable software within a research environment. | DeepMind | Coding ability alone does not establish research or safeguarding competence. |
| AI solution architect | Own solution design, governance, stakeholder decisions and delivery into live service. | Capgemini advertisement | The pasted Capgemini role is architecture, not a research-engineer job description. |
| MLOps engineer | Deploy, monitor, test and maintain model pipelines, data dependencies and releases. | Databricks careers | Strong DevOps crossover; not identical to training new models. |
| DevOps engineer | Automate build, test, deployment and infrastructure provisioning. | Canonical careers | Portfolio evidence: CI/CD, Terraform, containers and incident recovery. |
| Platform engineer | Create reusable infrastructure, developer tools and secure deployment foundations. | GitLab careers | Cloud architecture and automation overlap; individual team scope varies. |
| Site reliability engineer | Measure reliability, respond to incidents and improve operational resilience. | Cloudflare careers | Focus on observability, networks, troubleshooting and service quality. |
| Data engineer | Build reliable data ingestion, transformation, storage and governance workflows. | MongoDB careers | Supports AI systems, but differs from core model research. |
| Cloud engineer | Implement cloud networking, compute, identity and infrastructure. | Oracle careers | Cloud diagrams, infrastructure-as-code and permissions are useful evidence. |
| Cybersecurity engineer | Protect platforms, assess risk and implement monitoring and response. | Darktrace careers | Security engineering has distinct technical and governance responsibilities. |
| AI governance / assurance specialist | Evaluate AI risks, document safeguards and audit accountability. | Quantexa careers | Policy and technical skills overlap; qualifications vary by employer. |
| Technical programme manager | Coordinate dependencies, scope, evidence and stakeholders across teams. | Accenture careers | Transferable AV, project, operations and governance experience may help. |
| Technical support / service delivery manager | Own support quality, metrics, escalations, incidents and customer outcomes. | CGI careers | Connects troubleshooting, governance, stakeholder care and team leadership. |
“own the technical solution governance at all stages”
Accountability: requirements, deliverables, design documentation, architecture oversight and agreed specifications; work with agile delivery teams, senior stakeholders and vendors. The role can support commercial proposals.
Skills: solution architecture and generative AI; RAG, knowledge graphs, prompting, Azure OpenAI, agent frameworks, CrewAI and MCP. Python experience is advantageous.
Working pattern: hybrid across offices, client sites and home; the advertisement excludes fully home-based working.
Employer statements: inclusive culture, ongoing learning, training and wellbeing support. Disability Confident interview offer: declare a disability, meet minimum essential criteria and opt in.
Read the complete employer advertisement, benefits and company description. Employer promotional claims remain attributed to the employer; availability and terms require rechecking.
Correction: RAG means Retrieval-Augmented Generation. The Capgemini copy says “Resource”; the original paper supports the correct term. Lewis et al., NeurIPS 2020
Partial alignment: Python and AI architecture appear in the Capgemini role. Specific requirements for C++, C#, Rust, PyTorch, JAX, TensorFlow, LoRA, DPO, Ray or Slurm need evidence from an actual role; they are not universal eligibility requirements. Role-specific evidence
Not established: a general salary range, a guaranteed job outcome or a universal doctoral requirement. The pasted search overview and generic domain links do not substantiate those claims.
Design principle for this site: assess safeguarding competence independently of coding, architecture and research credentials. Approval needs evidence about protecting people, not only passing software tests.
Adult safeguarding guidance identifies empowerment, prevention, proportionality, protection, partnership and accountability. It concerns abuse and neglect and person-centred responses. DHSC — chapter 14
“Accountability and transparency in delivering safeguarding”
Suggested AI parameters: affected people, vulnerability, coercion and exploitation risks, consent boundaries, accessible reporting, trained human referral, response ownership and follow-up. Record domain and jurisdiction before choosing the applicable procedure.
ASVS provides requirements for application technical security controls. OWASP ASVS
Suggested evidence: authentication, authorisation, input handling, supply-chain checks, incident containment and tested recovery. A secure application can still cause interpersonal harm.
NIST AI RMF is voluntary guidance for incorporating trustworthiness into AI design, development, use and evaluation; its 1.0 framework is under revision. NIST — current framework status
Suggested evidence: contextual risks, validity, reliability, harmful outputs, fairness and human oversight. A benchmark score is not a safeguarding qualification.
England’s DfE product safety standards apply to generative AI in educational settings; the January 2026 update adds cognitive, emotional, social, mental-health and manipulation considerations. DfE — scope and update
Keeping Children Safe in Education is statutory guidance for schools and colleges in England. Adult safeguarding duties have their own scope and conditions in chapter 14 of the Care Act guidance. These are not universal duties imposed on every AI engineer. KCSIE · Adult safeguarding guidance
This is a proposed engineering evidence gate informed by the guidance above, not a certification or substitute for local safeguarding procedures. Do not enter real safeguarding cases into this public demonstration.
DSIT’s 2025 survey reports 35% of surveyed organisations struggling to fill AI roles; 97% of respondents reported at least one skills gap. DSIT / G&T executive summary
ONS estimates 702,000 vacancies for June–August 2026, down 1.1% on the quarter and 4.9% annually. ONS — September 2026 release
“2.5 unemployed people per vacancy”
DSIT survey respondents reported plans to adopt agentic AI within three years. DSIT / G&T survey
Anthropic’s Cadences report studies Claude usage and initial user-survey findings. Anthropic — provider research
Suggested interpretation: skills scarcity and a weaker overall vacancy market can coexist. Demonstrate practical experience and evaluate actual roles rather than treating an AI-growth narrative as a job guarantee.
Watch hypotheses: research-to-service reliability, efficient inference, agents with oversight, scientific AI and assurance. The sources above motivate these questions; they do not prove future hiring growth.
Counter-evidence: track removed or frozen roles, narrower eligibility, revised adoption plans and poor customer outcomes. Do not infer salaries, workforce demand or investment returns from fundraising headlines.
Suggested method: use a fixed basket of official role pages, deduplicate job IDs, record openings and closures, compare required skills and experience, and distinguish location, contract and seniority.
Record advertised pay only when the employer discloses it; separate base salary, currency, equity and benefits. Leave missing values unknown. Add official funding announcements, customer evidence and filings as separate signals, then revisit your interpretation.
Use ONS vacancies for the broad economy and DSIT survey for AI employer context. Different populations and periods cannot be combined into one supply–demand ratio.
50 priority companies to research and potentially approach across frontier AI, technology infrastructure, consulting, quantitative engineering and growth companies. This is a career-research shortlist, not a claim that all companies have suitable vacancies. Check each employer’s role requirements, UK or remote eligibility and recruitment status before applying.
| Company | Group | Suggested role lane | Suggested evidence angle | Source & verification |
|---|
Use the searchable table to shortlist by your role lane. Verify location, seniority, salary, application dates and ethical or mission fit on official employer pages. Prioritise roles supported by portfolio or prior-work evidence; track applications in the downloadable CSV. Fifty describes the database size, not live vacancies or submitted applications.
The growth-company filter includes Mistral, Cohere, Wayve, Isomorphic Labs, Holistic AI, Helsing, Synthesia and ElevenLabs. These are established growth-company watch targets, not a claim that they were newly founded in 2026.
Suggested watch criteria: founding and funding announcements, research releases, customer deployments, official hiring changes, runway disclosure and mission fit. Company career pages support identity and recruiting routes; they do not independently verify commercial success. Check sector-specific safety, privacy and ethical constraints before choosing an approach.
One publication at a time, with a source claim, short excerpt and a question to investigate. Types distinguish technical papers, vendor whitepapers and market surveys.
Loading publications…
Suggested workflow: move from market evidence to a deliberate research, career or commercial decision.
Choose one reproducible research question. Show a fair baseline, controlled changes, uncertainty and an honest contribution. Match that evidence to an actual role’s requirements, then explain its relevance in a short approach. Keep safeguarding evidence separate and identify specialist review needs.
Prepare a requirements map, architecture decisions, trust boundaries, costs, acceptance criteria and a live-service handover. Demonstrate stakeholder judgement and accountable governance alongside technical implementation.
Use a paper to propose a mechanism, then check adoption, customer value, costs and financial disclosures. Record evidence that would invalidate the thesis. Technical capability alone does not establish valuation or a trade.
I prioritise well-supported work, the problems it solves and the change it enables. I follow AI developments through primary research, email updates and professional networks, including LinkedIn, and publish at a sustainable pace as findings mature.
Source-supported facts have adjacent attributed links. Employer claims and provider research retain their origin; suggested portfolio actions and interpretations are labelled. A disagreement is qualified only as far as the evidence allows.
Role descriptions change by employer. Vacancy pages may close; this snapshot is not a live job feed. Quoted excerpts are intentionally short. The supplied search overview is a discovery lead, not independent proof of salary or demand. Recheck sources and standards before an operational decision.