Evidence-gated career operations pipeline
Directed and built with AI-agent assistance: immutable run artifacts, stable identifiers, employer firewalls, evidence checks, bounded queues, and a hard approval boundary around applications and outreach.
Selected work
A recruiter-readable portfolio of independent systems, software, and applied AI work I have designed or directed. AI-assisted projects are described as such; prototypes are not presented as production products.
Lead system / Agentic operations
I designed and operate a governed, semi-autonomous internal system centered on a persistent AI executive orchestrator and two dozen specialized worker roles. The architecture lets agents advance internal work without granting them authority to publish, spend, contact people, change credentials, or make commitments.
Written decision tiers and explicit human approval gates
Persistent context, scheduled executive cycles, and stable work state
Named owners, evidence-bearing artifacts, and exact read-back verification
Independent projects
Projects are ordered by relevance to agentic systems, automation, and operational reliability.
Directed and built with AI-agent assistance: immutable run artifacts, stable identifiers, employer firewalls, evidence checks, bounded queues, and a hard approval boundary around applications and outreach.
Designed a shared Vikunja operating layer with explicit ownership, inbox triage, priority bands, recurring work queues, deterministic safety checks, and proof-based closeout.
Built with Claude: a Python and TinyLlama exploration agent with persistent memory, visited-page controls, and recurring diary, essay, question, and letter outputs—without paid model APIs.
Directed and built with AI-agent assistance: a plain-text-backed private task interface and a local static prototype for starting one task with a calm, low-pressure countdown.
Defined the architecture and data model for a private media-production studio organized around durable assets, provider bindings, worker passes, and explicit generation and publishing gates. Implementation remains a future phase.
Operate an independent character-led AI media studio and automation lab spanning scripts, visual prompts, audio, video, written content, task management, and internal quality control.
Built with SimTheory: a private web chat prototype for Graham and Ny, supporting only Vietnamese and English and designed to adapt to their conversational patterns and preserve intended meaning rather than rely on word-for-word translation. Ny did not adopt it, and it was not deployed as a production product.