Below: Prepared briefings · scroll to access
Owning system design end to end on GCP Cloud Run — a 6-agent orchestration pipeline (Kopernica), the Mnemosyne memory subsystem, and a parallel voice sidecar that runs sentiment / query / analytics off the hot path. Currently shipping the production-grade Python + JavaScript SDKs and the multi-tenant BYOD surface that exposes it all.
Re-architected a serial 6-agent pipeline into a coordinator-driven DAG with prompt consolidation, model tiering, and async parallel execution. End-to-end latency 36s → 9.5s.
Context-scoped memory architecture with semantic retrieval and temporal-decay lifecycle (Core → Dream → Forgotten → Deleted), bucketed per user and domain so context cannot cross boundaries.
Parallel voice sidecar running sentiment, query, and analytics pipelines asynchronously off the hot path — they never add to perceived response time.
Redis-backed concurrency control with per-session sliding-window locks prevents state corruption when multiple real-time voice turns hit the same session simultaneously.
Shipped both SDKs covering orchestration, memory, analytics, sessions, and BYOD multi-tenant infrastructure — with CI/CD, audit logging, and signed webhook delivery.
CI/CD pipelines, audit logging, signed webhooks, circuit breakers, and graceful shutdown — all on Cloud Run + Cloud SQL + Memorystore behind a private VPC.
Four operational modules — focused capabilities backed by shipped production work.
Multi-agent orchestration with persistent memory and pipeline tuning.
Sub-200ms conversational AI with multimodal retrieval and streaming STT/TTS.
Event-driven systems with exactly-once semantics and observability.
Architecture reviews, AI roadmaps, and hands-on guidance for production teams.
A curated selection of recent production AI systems and full-stack work — each rendered as a JARVIS mission file.
Personal interest-graph and agentic recommendation engine turning cross-platform activity into a typed Neo4j interest graph with SSE-streamed recommendations.
AI meeting-intelligence platform automating pre-meeting research, live in-meeting assistance, and post-meeting synthesis with real-time voice and hybrid RAG.
Knowledge-graph retrieval engine that minimises LLM calls by doing entity resolution and validation in embedding and rule space; concepts live in hyperbolic (Poincaré) geometry.
