Cases
How the question changed.
How did I find a question we could act on? Each case starts with the reframing, then explains the approach, my contribution, and the result. Start with three representative cases; personal AI systems and prototypes follow below.
- DELIVERED
- ONGOING PRACTICE
- PROTOTYPE
Start here
Three cases that show how I think.
- Finance app: turning field records into issues a team can act onFrom organizing 254GB of records to finding grounds for product changes: review behavior by task and form 86 traceable issues.Delivery & progressDelivered 86 decision-ready issues in six categories, linked to tasks, observations, and priorities.EVIDENCE ARCHITECTURE
- Source recordsKeep tasks and context
- Tasks & observationsSeparate actions and interpretations
- Traceable issuesEach finding has a source
Records → Tasks & observations → Traceable issuesConceptual model
- Medical devices: testing while there is still time to change the designFrom completing a test to checking design while change is possible: connect critical tasks, possible use errors, and test timing to development.Delivery & progressHelped establish study plans, recording templates, and evaluation processes used in work for 25+ medical-device companies and institutions.DESIGN WITH FEEDBACK
- InsightDefine the assumption to examine
- Design ⇄ ObserveReturn feedback while changes are possible
- ConfirmationCheck whether the question was answered
Insight → Design & observation in a loop → ConfirmationConceptual model
- Service robots: trying the interaction before building the featureFrom waiting for a finished feature to testing understanding now: separate interaction stages and simulate key responses before further development.Delivery & progressUsed a four-stage model and Wizard-of-Oz studies, and helped document repeatable research procedures.PROTOTYPE / TWO LAYERS
- Frontstage · ExperienceStart, perform the task, end
- Backstage · SimulationA person follows response rules
- Observe & decideTest understanding before implementation
Observe understanding · Simulate the responseConceptual model
More project experience
Finding the right question in each setting.
- Automotive HMI: deciding what this round of research needs to answerConcept comparisons, pre-production checks, and field observation need different questions. I aligned the research with the development stage.Delivery & progressDeveloped an evaluation framework across four development stages for project delivery and research-team training.
- Dairy packaging: research that led to a usage instructionEEG, gaze, and observation each had a role. The task was to connect what people saw with how they picked up, opened, and used the packaging.Delivery & progressCombined measurements and natural-use observation to propose packaging changes, including instructions for using the fruit cup.
- Five Middle Eastern markets: what can be shared, what needs local workI combined desk research, interviews, surveys, and competitor analysis to distinguish shared needs, local conditions, and open questions.Delivery & progressOrganized cross-market needs and follow-up questions to support shared-product and localization decisions.
- Consulting: making project experience usable by the next personExperienced staff repeatedly had to start from scratch. I brought study design, responsibilities, and review into the delivery process.Delivery & progressHelped build an ongoing consulting practice, using project experience in proposals, team training, and knowledge retrieval.
AI work and experiments
I build tools for my own problems, too.
These include a system I use daily and prototypes awaiting validation with real research data. Each case states its operating status and what remains to be tested.
- My AI work system: connecting repeated lookup and admin tasksI use OpenClaw, Claude Code, and existing retrieval tools for email information, calendars, reference lookup, and filing, with defined write permissions.Delivery & progressUsed in my daily work, with task instructions, separate knowledge collections, and agent roles. No public efficiency benchmark is available.
- PsyPhiClaw: making multi-device analysis inspectable step by stepI split import, time alignment, analysis, and reporting into modules to explore reducing the preparation needed between research tools.Delivery & progressPublished 18 behavioral-analysis modules. Validation is paused; reliability on real research data remains unverified.
- BehaviorLens: keeping review points in automated video codingVideo coding takes time and requires context. I designed a three-pass prototype to explore where researchers should review automated work.Delivery & progressConcept, architecture, and frontend prototype prepared. The 20-minute processing goal and coding accuracy remain unvalidated.