UX leadership · AI and data solution delivery
My foundation is UX leadership across complex products and services. I’m expanding into AI and data solution delivery through implemented learning prototypes and ongoing Databricks work. The ten AI experiments below document methods and prototype results, with human accountability retained.
Expanding the practice
Hands-on work with explicit evidence limits. These projects show a developing delivery practice; production readiness and impact remain to be verified.
Implemented learning prototype
An AI-assisted support prototype with APIs, Postgres, agents and approval flows. My work includes problem definition, review, SQL checking and debugging.
Evidence limit: Deployment is unverified. No production outcomes are claimed.
Source prototype
A source prototype covering Python ingestion, SQL transformations, Parquet outputs and a Streamlit dashboard. The next step is to reproduce and check its results.
Evidence limit: Runtime is unverified. No operational performance or business impact is claimed.
In progress
Structured study and sample-data exercises are building the foundation for a Databricks pipeline. A completed pipeline or certification is not claimed.
Evidence limit: Work is in progress. No completed deployment or measured outcome is claimed.
Flagship workflow
A reconciliation loop connects coded components, tokens and Figma assets.
Code ↔ canvas
Specialised agents and critics turn structured requirements into a discussion-ready prototype.
Under 2 min
A fine-tuned 3B vision-language model performs repeatable first-pass interface critique locally.
Local 3B modelFrom evidence to evaluated experience
Capacity layer · Episodes 1 + 7
Automate repetitive production and coordination so senior talent can spend more time on decisions.
Evidence maturity
Most systems are working prototypes. Episode 09 has a prototype benchmark. Episode 01 has a reported task comparison (validation pending), not a verified result.
Complete evidence library
Repetitive production and coordination work consumes senior design capacity.
Reduce low-value effort and create more capacity for judgement, collaboration and quality.
Human control: Start with high-volume, low-risk tasks. Require approval before actions or outputs move forward.
An agent prepares a structured production output for human review.
Value: Reported task comparison: five days to thirty minutes for the defined task; validation pending.
Human: Human reviews the prepared output before use.
A personal control layer coordinates briefings, tasks and routine actions.
Value: Potential to reduce coordination overhead; team adoption requires access controls.
Human: User approves actions; organisational use needs access controls.
Research, customer reviews and behavioural data are slow to combine and interpret consistently.
Potentially compress evidence gathering, increase coverage and focus teams on journeys worth investigating.
Human control: Keep source links, privacy boundaries and researcher validation visible throughout the workflow.
Parallel agents discover, compare and organise sources for researcher validation.
Value: Potential to reduce evidence-gathering effort and broaden source coverage.
Human: Researcher validates sources, relevance and synthesis.
A LangGraph workflow structures customer reviews into themes and priorities.
Value: Potential to accelerate voice-of-customer analysis and issue detection.
Human: Researcher reviews themes against the source comments.
Seven agents connect GA4 and Microsoft Clarity signals to testable UX hypotheses.
Value: Potential to reduce analysis effort and focus investigation on priority journeys.
Human: Analyst validates hypotheses against journeys and context.
Ideas, requirements, design systems and prototypes lose time and meaning across handoffs.
Potentially shorten alignment cycles and reduce rework before formal design and engineering delivery.
Human control: Treat generated requirements and prototypes as drafts until product, design and engineering approval.A reconciliation loop connects coded components, tokens and Figma assets.
Value: Potential to reduce design-development drift and repetitive documentation.
Human: Design and engineering approve reconciled changes.
Agents challenge opportunities, surface assumptions and structure draft requirements.
Value: Potential to expose gaps earlier and improve discovery throughput.
Human: Product and design owners accept, reject or revise drafts.
Specialised agents and critics turn structured requirements into a discussion-ready prototype.
Value: A prototype was generated in under two minutes; design and engineering review remains mandatory.
Human: Design and engineering review every generated prototype.
Manual interface review is repetitive, specialist capacity is limited and some material is sensitive.
Potentially increase review coverage, surface drift earlier and protect sensitive interface material.
Human control: Use shadow mode, golden datasets, false-positive tracking and explicit human override.
Automated comparison flags likely interface changes for specialist review.
Value: Potential to detect defects and design-system drift earlier.
Human: Specialist confirms defects and release decisions.
A fine-tuned 3B vision-language model performs repeatable first-pass interface critique locally.
Value: Potential to reduce specialist pre-review effort while protecting sensitive material.
Human: Specialist owns the final critique and prioritisation.
Consulting adoption path