Opeoluwa Osho

An event-driven AI platform on Google Cloud that turns a list of product topics into finished, live ad campaigns — landing pages, Gemini-written copy and Gemini-generated visuals, an automated QA gate that blocks bad creatives before any ad spend, and launch to Meta with budget and bid control. Seven specialized services coordinated by a TypeScript orchestrator I built (~29k LOC): Cloud Tasks command queues out, Pub/Sub events back, with exactly-once processing (message-ID dedup) and a monotonic status state machine so duplicate or out-of-order events can't corrupt pipeline state. I architected the platform from scratch and personally built its core — the orchestrator and the React 19 operator console (~61k LOC authored solo, first commit to production in six months) — while a small team built the Python content service to the integration contract I defined; together we took a 2025 n8n prototype to this production system.

End-to-end RAG system for a finance client: thousands of pages of earnings reports, regulatory filings, and contracts, chunked structure-aware and embedded into pgvector — vectors and document metadata in one Postgres store, so similarity search and SQL filters run as a single sub-second query. A 5-person finance team asks questions in plain English and gets streamed answers grounded in retrieved passages, with every claim cited back to its source.

Go backend work on the text-to-image generation path of a major AI anime-art platform. Dynamic request batching (one GPU pass → multiple images), concurrent dispatch and worker pooling to keep RunPod / AWS SageMaker GPU workers saturated, warm model pools to kill cold starts, and pprof-driven hot-path tuning — more images per GPU-hour at lower per-image cost. Also integrated GLM-4, DeepSeek V3, and Claude for streaming chat across 105,000+ community models. The same backend powers the Moescape AI platform.