AI Product & Systems
I build production AI products and the platforms they run on.
The work includes restaurant intelligence, autonomous delivery, and household memory, along with the verification, cost, and governance problems those systems create once they leave a demo.
Data
semantic contracts
Agents
orchestration
Memory
context graph
Trust
evaluation loop
Field notes
Interactive demos
The archive examines evidence, verification, governance, memory, model routing, and replacement. The guide turns those essays into cited, practical answers.
Evidence-led AI
A public guide to the principles behind production AI systems. Answers draw from 23 essays, distinguish evidence from inference, and cite the supporting source.
Domain-specific example
Describe the domain and the decision it needs to support. A request can be shaped around its data, vocabulary, workflows, and evidence requirements.
Selected work
Product and architecture
SkyTab Intelligence is a restaurant product in which operators ask questions in ordinary language and receive reports, dashboards, and recommended actions. Every displayed metric has to reconcile to the source system, and the query that produced it travels with the result so the number can be verified for accuracy.
The platform is a composable multi-agent system with penny-exact metric reconciliation. Benchmarks belong to the merchant, using that restaurant's history, seasonality, and goals rather than an industry average.
Enterprise search
Merchant Explorer is a find-anything architecture for merchant data. It combines ontology discovery, structured search, aggregation, and semantic search so the model can explore large object schemas without a hardcoded field list.
The model discovers and queries objects instead of memorizing the schema. That is how search survives a catalog that would overflow context if the field list were dumped into the prompt.
No-code agent platform
Enterprise Agent Builder is a visual canvas for assembling agents: connect data, define context, choose a reasoning topology, attach tools, and publish a headless MCP agent. It uses the same capability set developed for Merchant Explorer.
Non-engineers can assemble an agent. They cannot bypass access, tools, or context policy, because those checks live in the compile path rather than on the canvas.
Autonomous development agent factory
The Dark Factory is an autonomous development loop in which agents plan, build, verify, triage, and fix through gated phases. The worker that changes the code can report success, but it cannot accept the change.
One 49-hour run completed 124 iterations and 40 fixes, then stopped at 78.85 against a release threshold of 80. Activity was not treated as completion.
Family operating system
Lucky and Clover are two household agents that share family memory, member context, routines, and home signals. They are available by voice or chat.
They coordinate chores, school, shopping, schedules, and the connected home from one shared context. A memory can be reviewed, shared, and withdrawn on the next turn rather than overnight.
Agent memory architecture
Memory Graph stores facts, preferences, relationships, and prior interactions with ownership, sharing, review, and revocation. Retrieval is not enough; access has to be enforceable on the next turn.
Memories are learned in conversation, reviewed by the user, scoped, and connected in a graph. A share can be withdrawn mid-session, and the next turn has to behave as if the fact is gone.
Voice transcription and multi-voice hub
MLX-Voice is a local workspace for transcription, capture, and multi-voice work on Apple Silicon. Latency is measured at the boundary a listener actually hears, not only at the warm-model microbenchmark.
A warm text-to-speech microbenchmark hit 0.19 seconds, while the live path to first audio was about nine seconds. Selection and optimization should use the second number, because that is the latency a listener experiences.
Enterprise AI enablement
A library of live trainings, example skills, and micro-projects for product, risk, compliance, QA, legal, and commercial teams. The curriculum is aimed at production failure modes rather than prompt technique alone.
The sessions were delivered live with the people who own the workflow. The material covers context, contracts, and verification, not a catalog of clever prompts.
Workflow transformation
The SOP Builder turns messy procedures into playbooks, agent instructions, checklists, and workflows. Policy, steps, and approvals remain attached to the result.
The procedure has to exist as a contract before an agent can run it. Otherwise the organization has automation without an owner.
Agentic risk operations
An AI coworker for credit risk analysts that reads financial statements, calculates exposure, drafts review memos, and generates leadership summaries. The work sits inside an underwriting workflow rather than a generic document chat.
The system was built for periodic review and portfolio monitoring. The memo has to survive an audit, not only a first reading.
Field notes
Deterministic tests, behavioral evals, traces, and hardening loops exist so a bad answer, a bad metric, or a bad merge is visible in time to stop it. A benchmark that cannot be rerun against the current artifact is a story, not a control.
SkyTab Intelligence became more useful when the comparison set belonged to the merchant: local history, local seasonality, and local goals. Generic peer numbers flatten the decision. Tenant-local inference is part of the product, not a modeling flourish.
The interesting part of an autonomous loop is who owns acceptance and how findings get classified, not the novelty of an agent writing code. Planning, gates, drift checks, review, and triage are the management system. A loop that cannot fail after a confident completion report is unsupervised typing.
Family and enterprise memory both need review, scope, recalibration, and graph structure. Recall becomes a product surface when a user can correct, share, and withdraw what the system thinks it knows, and when that withdrawal takes effect on the next turn.
Archive material