AI Product & Systems
A working archive of production AI products, agentic platforms, and operating systems.
Product notes, systems work, and practical AI research across agentic platforms, semantic layers, memory, and enterprise adoption.
Data
semantic contracts
Agents
orchestration
Memory
context graph
Trust
evaluation loop
Latest writing
Transition patternTransition pattern
Major technologies are misread at the start, adopted unevenly, governed late, and only pay off once work is redesigned around them. AI is running the same arc, faster. Treat it as an operating-model transition, not a tool rollout.
Productivity lagProductivity lag
When factories first electrified, they bolted motors into steam-era layouts and saw little. The gains came only after the work was redesigned. Expect the same gap between AI access and AI productivity, and respond with disciplined learning rather than waiting.
Democratized powerDemocratized power
VisiCalc and Lotus 1-2-3 did not just speed up arithmetic. They changed who could model a business, and they buried assumptions inside formulas. AI democratizes analytical production the same way, and makes flawed reasoning look fluent. The answer is review that scales with the new producers.
Governance lagGovernance lag
In 1995, 14 percent of US adults were online and famous voices predicted collapse. They were wrong about diffusion and often right about the institutional problems uncontrolled adoption would create. Separate the two questions for AI: will it spread, and what controls does safe use require?
Common patternsCommon patterns
Across electricity, spreadsheets, computers, and the internet, the same six organizational dynamics decide whether a technology becomes an advantage or unmanaged risk. Here they are, and what each one demands of anyone preparing for AI.
Jagged frontierJagged frontier
AI is a general-purpose technology whose value depends on complementary innovation, just like electricity and the internet. But it differs in three ways that change how you manage it: it targets high-skill work, its barrier to entry is language, and it is brilliant and wrong at lookalike tasks.
Operating playbookOperating playbook
If AI is an operating-model transition rather than a tool rollout, preparation is concrete work. Here are ten steps, from treating the task as the unit of analysis to measuring outcomes instead of activity, that turn experimentation into governed advantage.
Three horizonsThree horizons
Manage the AI transition in three overlapping horizons: controlled enablement, workflow redesign, and business-model transformation. It is the path from safe access to redesigned work to new value, and it avoids both passivity and reckless acceleration.
Cost per taskCost per task
Do not buy model hype. Benchmark the work your company actually does, then route each job to the cheapest model that reliably clears the bar.
Inference controlInference control
Inference is becoming strategic infrastructure. The question is not only which API is cheapest, but which parts of the AI stack your company controls.
AI operating modelAI operating model
Enterprises need an internal AI routing layer, not only access to ChatGPT, Claude, Gemini, Cursor, or whatever lab surface happens to be most popular this month.
Planning systemsPlanning systems
I do not ask AI to decide. I use it to make ambiguity visible before execution starts.
Signal qualitySignal quality
A feed audit showed the real AI content problem: not that everyone writes about AI, but that everyone writes about it the same way.
Enterprise adoptionEnterprise adoption
Most adoption programs fail because they treat AI literacy as one skill. It is really a progression from answers, to collaboration, to orchestration.
Organization designOrganization design
As AI commoditizes knowledge, the scarce capability becomes designing, directing, and validating systems of work.
Enterprise architectureEnterprise architecture
In an AI market that changes monthly, durable advantage comes from systems designed to be replaced.
Selected work
Product and architecture
A production restaurant intelligence product built from scratch: conversational reporting, dashboards, semantic metrics, typed artifacts, and source-backed insights.
Built as a composable multi-agent platform with penny-exact metric reconciliation and merchant-specific inference, so every benchmark is local, relevant, and able to improve over time.
Enterprise search
A find-anything architecture for merchant data using ontology discovery, structured search, aggregation, semantic search, and iterative schema exploration.
Designed to avoid hardcoded field lists and scale across large object schemas without flooding the model context.
No-code agent platform
A visual builder for enterprise agents: connect data, define context, choose reasoning topology, attach tool skills, and publish the result as an accessible headless MCP agent.
Designed so non-engineers can assemble governed agents using the same reusable capability set developed for Merchant Explorer.
Autonomous development agent factory
A structured autonomous development lifecycle where agents plan, build, verify, triage, fix, and ship through gated phases.
Evolved into a repeatable operating model for high-throughput AI-assisted engineering.
Family operating system
A family AI platform where household memory, member context, routines, and home signals combine into one shared operating layer.
Lucky and Clover act as two coordinating agents available through voice or chat, using whole-family context to help with chores, school, shopping, schedules, and connected-home routines.
Agent memory architecture
A memory model that turns onboarding, recalibration, preferences, facts, relationships, and prior interactions into useful future context.
Built around continuity, permission, and practical recall: memories are learned conversationally, reviewed by the user, and connected through graph structure over time.
Voice transcription and multi-voice hub
A local voice workspace for transcription, voice capture, and multi-voice workflows built around Apple Silicon and practical operator use.
Extends AI interaction beyond text into fast local voice workflows and reusable voice infrastructure.
Enterprise AI enablement
A large body of trainings, example skills, prompt patterns, and micro-projects used to help teams adopt AI tooling responsibly.
Delivered repeated live training with practical examples for product, risk, compliance, QA, legal, and commercial teams.
Workflow transformation
A system for turning messy procedures into structured operating playbooks, agent instructions, checklists, and reusable workflows.
Targets the unglamorous but valuable enterprise layer where AI needs policy, steps, approvals, and durable documentation.
Agentic risk operations
An AI coworker for credit risk analysts that reads financial statements, calculates exposure, drafts review memos, and generates leadership summaries.
Designed around auditable workflows for underwriting and portfolio monitoring rather than generic document chat.
Field notes
Layer deterministic tests, behavioral evals, trace review, and repeated hardening loops. The goal is not a permanent benchmark claim; it is a system that makes failures visible and repairable.
Restaurant benchmarks become more useful when they belong to the merchant: local history, local seasonality, local goals, and inference that learns from the actual operating context.
Autonomous development works best as a lifecycle with planning, gates, drift checks, review, and triage. The interesting part is the operating cadence, not the novelty of one agent writing code.
Family and enterprise memory both need review, scope, recalibration, and graph structure. Recall becomes a product surface when users can correct what the system thinks it knows.
Archive material