Giorgi Giorgobiani

• Architected an AI companion platform end to end in Rust (tokio, axum, tonic): custom ESP32-S3 hardware, a three-tier memory system, and a multi-tenant cloud backend licensed as a white-label platform to third-party brands.
• Built three memory tiers queried in parallel every turn — Redis for session state and hot facts (<5ms), Qdrant for episodic, semantic and emotional memory ranked by similarity (<50ms), and an Amazon Neptune knowledge graph for entities and relationships (<150ms). Episodic memories decay on a 30/90/180-day schedule by domain; anything accessed repeatedly and high-confidence is promoted into the permanent graph.
• Built a streaming voice pipeline (streaming STT → LLM → custom TTS) tuned to a 520ms latency target against an 800ms hard budget, with self-hosted LLM, embedding and emotion-model inference on AWS SageMaker (Inferentia/Trainium) behind a trait-based provider abstraction with multi-AZ failover.
• Enforced per-child memory isolation across licensed brands with a separate Qdrant collection and a separate Neptune subgraph per child. Internal services communicate over gRPC/tonic, with WebSocket binary framing for the real-time audio protocol and OpenTelemetry/X-Ray tracing across the critical path.

• Architected a multi-tenant conversation-intelligence platform converting ~100 hours of calls a week into structured business intelligence for several teams.
• Built a 9-worker, queue-driven pipeline across 7 queues with 7 dead-letter queues, bounded retries, and a shared concurrent message-pool library as the reliability backbone.
• Designed a 4-stage signal promotion ladder that never demotes, lifting per-call learnings into cross-call signals and 31 prerequisite-chained artifact types evaluated in 2 database queries.
• Built retrieval over 3,072-dimension embeddings in pgvector across transcript segments, summaries and reflection insights.
• Shipped a chat assistant with a 24-tool function-calling registry using thinking-token streaming and parallel tool dispatch over SSE, plus an MCP server exposing the intelligence layer to external agents.
• Guaranteed tenant isolation via PostgreSQL row-level security over a 50+ model Prisma schema in a single-deployment architecture, with all infrastructure as code at full dev/prod parity.
• Cut compute cost ~94% by moving off per-task serverless containers to 3 EC2 capacity providers tuned per workload shape, autoscaled on queue backlog per task rather than CPU.

• Built the NLP vertical on a decentralized data platform: a 10-agent pipeline where a meta-orchestrator fans batched events out to 7 model-backed agents and 3 coordination agents in parallel, merging into a social graph through a single writer.
• Implemented content-hash idempotency making the pipeline fully replayable — Postgres can be wiped and rebuilt from source events with identical output.
• Built semantic retrieval over 384-dimension embeddings in pgvector with dynamic topic clustering at a 70% similarity threshold, over a 27-table schema with 49 indexes.
• Exposed 5 query operations as MCP tools through the platform gateway, making the social graph directly queryable by external LLMs.
• Declared pipelines as infrastructure-as-code (CDKTF, custom resource types), so a new pipeline ships as a stack file and deploys without recompiling any backend service.
• Integrated the NLP vertical against a Substrate-based chain (five custom runtime pallets) and a signed-event ingestion pipeline — ed25519/sr25519 client signing, content-addressed storage — consuming from the same Kafka Streams topics and RocksDB-backed state store the platform's stream-ETL layer runs on.

Also read and integrated directly against, without owning: Go (the platform's unified ingestion/compute engine), Kotlin (Quarkus webhook service, Kafka Streams ETL topology).

• Founding engineer at a legal-AI startup; built the full platform from zero (AWS, Next.js, Node.js, TypeScript), serving 2,000 users.
• Built the document-intelligence pipeline behind contract due diligence: parsing and chunking, embeddings, and LLM-driven extraction, cutting the manual review hours attorneys spent per deal.
• Orchestrated that pipeline as coordinated Lambdas with SQS queueing, offloading heavier LLM processing to EC2 where Lambda limits bound throughput.
• Designed AppSync GraphQL endpoints for real-time updates and offline sync, cutting API response times 40%.
• Led the company through SOC 2 Type 2 certification.

• Developed prompt strategies and evaluation loops for LLM output quality, iterating on model behaviour against project-defined objectives.

• Built a model-federation dashboard letting researchers schedule, sequence and automate model-training jobs, removing most of the manual coordination from the training workflow.
• Visualised the full training-node topology and inter-job relationships as an interactive browser graph over Neo4j.
• Built full-stack applications (React/Redux, Java/Vert.x, Node.js, GraphQL) deployed on AWS with Docker and Kubernetes.

• Built a Ruby proxy layer for frontend request capture and third-party integration, and led the team's adoption of React and Node.js.

• Contributed to a monolith-to-microservices migration (Node.js/Express, Ruby on Rails), built React/Redux/TypeScript frontends, and deployed serverless services on AWS Lambda.