Four years of compounding practice
I have worked deeply with generative AI since 2022—not as a layer added at the end, but as a new material for products, teams, and operating systems.
My work spans productized personalized feedback, adaptive roleplay, AI-readable design systems, agent teams, production delivery, evaluation systems, local models, dataset development, and fine-tuning open-weight models.
4years building with generative AI
End to endproduct strategy, design, engineering, and model systems
Local + cloudthe runtime follows the problem and the data boundary
00 The thesis
The model is only one part of the work. Dependable AI needs structured context, bounded jobs, observable behavior, explicit quality standards, and feedback that improves the next run.
01Product systemsExperiences people use
02Operating modelsSystems teams can scale
03R&D infrastructureTools for what comes next
01 Design systems + craft
I used AI to stand up design-system foundations quickly, then shaped those foundations into a shared language for people and agents. That made stronger product building available to more of the organization while giving me more time to push interaction quality, visual detail, edge cases, and the final implementation much further.
01Branded componentsTokens, typography, spacing, states
02StorybookInspectable patterns and behavior
03Product contextStandards, APIs, and domain language
04Quality gatesAccessibility, localization, linting, review
Builders and specialist agents started from real components, explicit standards, and known service contracts instead of a blank canvas. Coherent output became less dependent on knowing every unwritten rule.
What changed Shared patterns replaced one-off invention, relevant context moved to the point of work, and ideas reached working software faster.
AI compressed the mechanical implementation loop. I could design directly in working software, test with real behavior and data, and spend more of the cycle refining hierarchy, transitions, responsive behavior, states, and the details that make a product feel considered.
What changed More meaningful iterations happened before release, refinement moved into production code, and craft was judged in context.
Foundation test5 distinct product experimentsCompared component foundations against real product problems before choosing.
System depthBeyond default componentsRebuilt the visual language through branded tokens, typography, states, and interaction patterns.
Organizational resultTest space → production platformThe rapid experiment environment became the main place customer-facing product work shipped.
02 The full practice
Prompty and evals are part of the picture. The broader work connects products, agent orchestration, model infrastructure, learning systems, and organizational change.
I design the work before, during, and after the model, then give people and agents a shared design language that keeps the result coherent.
Dais + Cairn Adaptive feedback and structural intelligence
BENlabs system Branded components, standards, and API context
Ralph-style loops, specialist roles, planning records, quality gates, and explicit human decisions turn working proof into a repeatable operating model.
Planning Harness + Prompty Shared plans, deployable loops, and guardrails
Builder portal AI-native experiments became production
Challenge cases, annotations, deterministic checks, local inference, and fine-tuning infrastructure turn judgment into an inspectable system that compounds.
Prompty Proof Comparable runs and regression evidence
Witness + training pipeline Local models, annotations, and dataset export
03 Productized intelligence
I have productized models across text, image, voice, video, and 3D workflows. Sometimes one model is enough. Often the value comes from composing several stages, choosing the right runtime, and designing the experience around the result.
The architecture changes. The product standard stays visible.
01Personalized feedback
02Working software
03Generated media
043D + video assets
04 The connective tissue
Give the work shape, context, and a bounded objective.
Choose the model, tools, roles, permissions, and handoffs.
Preserve outputs, behavior, provenance, and runtime evidence.
Compare the result against explicit checks and human judgment.
Improve the product, workflow, dataset, or model.
Personalized feedback conversation → transcript + signals → tailored debrief → next attempt
Model training generated UI → critique → designer annotation → training dataset → fine-tune
Agent teams scoped goal → specialist work → quality gates → human decision → reusable context
05 The compounding arc
Moved beyond isolated prompting into code, structured workflows, and early agent experiments.
Used AI-assisted development to build an independent React Native proof in four weeks and change what the organization believed was possible.
Built the design and frontend system around context, specialist agents, automated review, and working software.
Labsly spans adaptive feedback, agent workflows, local inference, datasets, evaluation, and open-weight fine-tuning.
06 What stays constant
Good context, clear roles, and explicit boundaries do more than an elaborate one-off prompt.
Outputs become dependable when cases, checks, provenance, and human review are designed into the system.
Privacy, latency, cost, and control are product decisions—not infrastructure footnotes.
AI can increase leverage without hiding who set the standard, approved the result, or owns the decision.