Portfolio
AI Platform Architecture Portfolio
A collection of personal lab and open-source projects for AI platform, robotics cloud, agent engineering, and distributed systems roles.
Production-related content only covers systems I actually owned, fully sanitized. Lab projects are never presented as production scale. Articles may be AI-assisted, but all code, tests, and failure reproductions are verified by me.
Links: GitHub · AI Video Studio · Margrop Labs · Coding Plan Dashboard
Selected Engineering Work
01 · Robotics Cloud Platform to AI Platform
Long-term owner of the robotics cloud platform’s OpenAPI, device state, task scheduling, async processing, idempotency, recovery, and production observability. Mapped onto a companion-robot AI Platform, the focus is organizing device identity, session state, tool invocation, model routing, long-term memory, and failure recovery into a unified platform boundary.
Open to discuss: multi-tenancy, device/family-member isolation, state consistency, reconnection, permissions, and confirmation for high-risk actions.
02 · AI Video Studio
A provider-neutral AI content pipeline covering Planner, Shot, Queue, Worker, Provider, Storage, and Render.
Open to discuss: task idempotency, worker leases, timeout/retry, provider rate limiting, async TTS/video tasks, cost accounting, MCP tool boundaries, and observability.
03 · Multi-Agent Shared Memory
Exploring shared context, memory extraction, storage, retrieval, conflicts, and isolation across agents — verifying how different agent frameworks collaborate through unified WebSocket/HTTP/database boundaries.
Open to discuss: memory freshness, correcting wrong memories, family-member isolation, sensitive-field masking, degradation, and rollback.
04 · AI Gateway / Quota / Cost
A self-hosted dashboard with provider adapters that unifies visibility into model quotas, costs, request status, and failure reasons.
Boundaries: only credentials I am legally authorized to use; no cookie storage, no bypassing server-side limits, no scraping others’ sessions; secrets come only from environment variables or a secret store.
05 · Observability and Recovery
Connecting logs, metrics, traces, event timelines, capacity, and incident reviews into an evidence chain — instead of showing a single “service is fine” screenshot.
Open to discuss: P95/P99 latency, retry storms, circuit breaking/degradation, message backlog, worker crash recovery, alert noise, and capacity management.
How I Work
- Define boundaries first: identity, permissions, data, failure modes, and rollback paths.
- Then platformize: provider abstraction, unified contracts, idempotency, rate limiting, degradation, and version governance.
- Finally close the evidence loop: logs, metrics, tracing, tests, capacity, cost, and incident reviews.
Last updated: August 21, 2026 · Personal portfolio, not an employer statement.