The newest platform in the range
I build on AI. I don't just use it.
AI is the newest platform I architect on, and I treat it like any platform: build on it, evaluate it honestly, know when to decline it. The Surveyor Agent and Inspector Agent are custom-built reasoning agents in production.
Deep on Microsoft Copilot and Claude, with Gemini in the working rotation and OpenAI's image models driving the generative pipelines.
Selected AI-engineering work · sheet index
Anthropic ClaudeMCP · agent designCDP
agent layer · reasoning modeled from senior-architect judgment
The Surveyor Agent - a stack-diagnostic reasoning agent
- Problem
- Platform audits happen after scoping; by the time anyone reads the configuration, the misconfigurations are already load-bearing.
- Read
- Catching broken configurations is pattern recognition developed across years of ecosystem audits — a consistent heuristic, and a consistent heuristic can be modeled into a diagnostic layer.
- Call
- Architected and shipped the Surveyor Agent: a diagnostic reasoning layer that validates stack configuration integrity before scoping. Proven against a hard real-world case — a multi-hop configuration relay no one owned end to end — audited front to back, with the gaps surfaced.
- Outcome
- In production at codedrovers.com. Its forerunner — an earlier discovery agent I built solo — was showcased at the Bloomreach American Town Hall: one of a handful featured worldwide, the only solo entry among teams of seven to nine, and a standout in a field of ~250 across 35 partner agencies.
estate auditverificationdeployed
every file claims to be canon → the one the bytes actually back
How I know which "source of truth" is actually true
- Problem
- In a heavily-iterated system, every file claims to be the source of truth, and they can't all be right.
- Read
- You can't trust what a document says about itself. The reliable signal is the bytes: what's current, what's a stale copy, what quietly contradicts what.
- Call
- Built and deployed an estate-audit layer that reads contents, not labels: it names the version the evidence supports as canonical, flags what's drifted, and abstains with receipts when the bytes genuinely don't decide. Adversarial "cry-wolf" cases gate every verdict before it ships; a wrong call gets reopened against the record and corrected, never overwritten.
- Outcome
- An estate you can trust: every "this is canon" claim checked against what's physically there, with the evidence attached to the verdict.
Pythonprivacy engineeringdeterministic
raw content → content-free flags → agentic triage
A privacy boundary the model never crosses
- Problem
- To let an AI audit a system, you usually have to show it the system - including the parts that must never leave the building.
- Read
- The agent doesn't need the sensitive value. It needs to know a value is there and what kind. Those are different, and the difference is the whole safety case.
- Call
- Engineered a deterministic scout that inspects the cells and emits content-free flags (never a raw value or a matched string): it sees the cells so the agent only ever sees flags, and nothing sensitive crosses the model boundary.
- Outcome
- An AI that can reason about sensitive data without ever being handed it. The boundary is the feature.
PythonModel Context Protocolopen source
a deterministic audit engine → a public MCP server, shipped
An open-source MCP server that hands an AI judgment, not just a diff
- Problem
- Everyone with a drafts pile has the same one: a dozen near-identical files, no reliable way to know which is current, and every dedup tool happy to delete the wrong one.
- Read
- A diff tells you what changed and refuses to say whether it matters. The missing piece is judgment, read from the bytes rather than the filename, and a model can supply it if you hand it the right deterministic evidence.
- Call
- Built and shipped a public MCP server (Python) that gives any AI assistant that evidence: which lines exist in only one place, which file is a subset of another, what actually changed, plus the rules to act on it. Local-first, no API key, diagnose-only; it proposes, it never deletes.
- Outcome
- Live and open-source at github.com/kellyorkin/cane-mcp. Anyone can install it and get an evidence-backed answer in one command.
also in the drawer
- AI-05
- Generative content pipeline at collection scale — Python + image-model pipelines generating to a schema'd JSON catalog; 100+ artifacts served live with CI/CD deployment across my own domains.
- AI-06
- Production MCP server, both sides of the protocol — auth-gated TypeScript server on the Anthropic Messages API, with client-side tool-use loops implemented directly against the wire format.
- AI-07
- Multi-model ensemble, on the record — Claude, Copilot, and Gemini under scoped roles, written compacts, ADR decision records, and source-provenance rules.