I build controlled automation systems that turn product ideas into testable, operator-friendly software. My focus is the hard part around AI features: deterministic state, review workflows, idempotency, budget controls, safe API boundaries, and recovery when an external result is uncertain.

AI agents contribute to implementation; evidence, approval, and release authority remain human-owned.

• Python control planes for multi-stage automation.
• Human-in-the-loop AI workflows with explicit approval gates.
• API integrations designed for retries, ambiguity, and partial failure.
• Content and media pipelines with provenance and quality validation.
• Operator interfaces that separate verified facts from unverified claims.
• File-backed agent workflows with durable handoffs and deterministic checks.

A Python control plane for short-form content operations: brief intake, AI-assisted scenario review, paid-generation approval, media inspection, deterministic package assembly, and separately gated publication workflows.

Key engineering work:

• append-only JSONL journals and replayable state machines;
• idempotent callbacks, jobs, receipts, and recovery paths;
• separate approval boundaries for generation, spending, and publication;
• provenance-bound media intake and FFmpeg/FFprobe validation;
• guarded integrations for Telegram and social platforms;
• offline-first verification with more than 1,200 automated tests.

A modular vertical-video pipeline covering scenario drafting, human review, quality gates, dry runs, generation ledgers, channel packaging, and upload readiness. Mock providers and non-paid status commands are the default; live generation and publishing require separate authorization.

A practical workflow for coordinating Codex, Claude Code, and OpenCode-style executors without turning them into autonomous release authorities.

It includes:

• narrow task packets and explicit executor boundaries;
• durable file-backed handoffs and run evidence;
• context selection, compression, and degradation checks;
• deterministic evaluation gates and guarded LLM-as-judge experiments;
• safe model routing, approval policies, and recovery-oriented tooling.

A Next.js, React, and TypeScript dashboard for comparing AI inference services. The product model distinguishes public evidence, provider-reported claims, and unverified data instead of presenting time-sensitive availability or pricing as permanent fact.

Python operator bots for review and approval workflows, with owner checks, idempotent callback handling, guarded downstream actions, safe error redaction, and explicit boundaries around paid or public operations.

A vertical-media pipeline explored through scenario planning, asset generation, rendering, packaging, and channel operations. This work informed the stricter state, provenance, approval, and recovery architecture used in newer projects.

Additional work covers Web3/DeFi integrations and decentralized AI-compute infrastructure. Public artifacts currently include a Gonka AI Compute fork and static developer-compliance pages.

flowchart LR
    A[Product intent] --> B[Deterministic plan]
    B --> C[Bounded AI or API task]
    C --> D[Evidence and validation]
    D --> E{Human approval}
    E -->|Revise| B
    E -->|Approved| F[External action]
    F --> G[Receipt and recovery state]

The default design principles are:

offline-first · fail-closed · idempotent · append-only · evidence-led · budget-aware · recovery-ready

I am turning the strongest private project evidence into small, sanitized, reproducible public editions. Until those editions are ready, project summaries describe verified engineering work without exposing private source, credentials, runtime data, internal infrastructure, or personal information.