Joseph Caughey
ARTIFICIAL INTELLIGENCE
Agents that build. Validation that proves they got it right.
Most teams ship AI-generated code and hope. The interesting work was never the generation, it is the proof: a spec the agent is measured against, evaluators that grade every seam, and validators that keep running once it is live. Trust is not a feature you bolt on afterwards.
What I Build
AI Capabilities
Deterministic Validation
Executable specs, contract tests and LLM-as-judge evaluators that grade generated work against a definition of correct rather than a vibe.
Agentic Engineering
Agents that write and modify production code inside defined permissions, scope and review gates, instead of being handed a repository and trusted.
MCP Tooling
Model Context Protocol servers that give agents real access to Jira, Figma and Git, so they act on the actual system rather than a description of it.
Workflow Automation
Design-to-code pipelines and ticket automation that remove the manual translation steps between what was decided and what gets built.
LLM Engineering
Structured outputs, tool calling, retrieval and evaluation harnesses, built for production workloads rather than demos.
AI-Generated Media
Film, music and imagery produced end to end with generative tooling, directed rather than prompted once and posted.
In production
Where this actually runs
Parity Integration
The validation framework behind production rebuilds, plus the Figma-to-code, Jira and MCP tooling around it.
Open sourceDirectiveAI
A Git-first standard that compiles AI directives into scoped, policy-gated, auditable branches.
ArchitectureContract-as-Code
Collective agreements compiled into executable compliance rules, then used to validate payroll calculations against them.
Credential
Certificate in Applied Artificial Intelligence, University of Victoria, completed with a 97% standing. Verify
The loop
Spec. Build. Validate. Ship.
Every feature starts as an executable specification rather than a prompt. Agents generate against that spec, inside defined permissions and scope. Evaluators and contract tests grade the result at every seam before it reaches main. Then the validators keep running after deploy, so a regression surfaces in minutes instead of in a support ticket.
It is an unglamorous answer to the only question that matters once AI is writing production code: how do you know it is right?
Work With MeTools & Frameworks
The Stack
Arranged by where each thing does its work. A wall of logos tells you nothing about how a system is built.
-
Spec
What correct means, written down before anything is generated.
- Executable specs
- Figma to code
- Jira
-
Build
Agents generate against that spec, inside defined scope.
- Claude Code
- GitHub Copilot
- Claude
- OpenAI
- MCP servers
- Scoped permissions
-
Validate
Every seam graded before anything reaches main.
- LLM-as-judge
- Contract tests
- Eval harnesses
- Structured outputs
-
Ship
Deployed, and still being measured.
- GitHub Actions
- Docker
- AWS
- Firebase
- Node.js
Models & media
Made with it
