New York, NY · Remote (US)

Giorgi Giorgobiani

I build the infrastructure that agent systems run on — orchestration, retrieval, evaluation, and the cost and latency controls that decide whether a system survives contact with production.

Nine years of production engineering, the last several on LLM and agent platforms.

See the workGitHubLinkedInRésumé

Selected work

Four systems

Senior Platform Engineer · Nov 2025 – Jul 2026Multi-tenant conversation intelligenceVenture teams sat through roughly 120 calls a week — about 100 hours. I architected the platform that turned that stream into structured, queryable business intelligence for several teams at once. Recordings land in a nine-worker pipeline across seven queues, each with its own dead-letter queue and bounded retries. Retrieval runs over 3,072-dimension embeddings in pgvector; on top sits a chat assistant with a 24-tool function-calling registry and an MCP server that exposes the intelligence layer to external agents. Tenants are isolated by Postgres row-level security in a single deployment rather than one deployment per customer. Compute cost fell about 94% when I moved off per-task serverless containers to three EC2 capacity providers tuned per workload shape, autoscaled on queue backlog per task rather than CPU. 100 hrsingested weekly9workers7queues + DLQs3,072dim pgvector24assistant tools~94%compute cost cutRead the teardown →
Senior Software Engineer · Apr 2025 – Oct 2025A ten-agent pipeline that can be replayedA firehose of chat events needed to become a queryable social graph — users, messages, topics, inferred relationships — without ten agents corrupting shared state on the way in. A meta-orchestrator fans each batch out to seven stateless model-backed agents in parallel; three coordination agents handle the rest, and a single writer merges every result into the graph. Content-hash idempotency at the storage layer made the whole pipeline replayable: wipe Postgres entirely, rebuild from source events, get the same graph. Retrieval runs on 384-dimension embeddings with topic clustering at a 70% similarity threshold over a 27-table schema. Five query operations are exposed as MCP tools, and pipelines themselves are declared as infrastructure-as-code, so a new one ships as a stack file without recompiling any service. 10agents per batch7 → 1parallel, single writer384dim embeddings70%cluster threshold27table schema5MCP query toolsRead the teardown →
Senior Software Engineer · Apr 2025 – Oct 2025A blockchain that only stores who, not whatA Substrate-based chain paired with an off-chain, content-addressed storage layer: a permissioned network of staked storage and CDN nodes holding the actual data, with the chain itself holding only registry state — node identities, clusters, customer accounts, staking, governance — across five custom runtime pallets. I built the NLP vertical on top of this platform (a separate case study on this site); the chain, the pallets, the ingestion gateway, and the query gateway were built by the protocol team. What follows is the platform substrate I integrated against closely enough to draw accurately: a signed-event ingestion pipeline where every client signs with an on-chain key before anything is admitted, a Kafka Streams topology that chains events into a content-addressed DAG with exactly-once semantics, and — the pattern I find most interesting in the whole stack — pipelines declared as infrastructure-as-code over four custom resource types, so a new vertical ships as a Terraform stack file and is exposed to LLMs as MCP tools automatically, without a single line of platform code changing. 5custom runtime pallets15sub-repos, one platform4IaC resource typesed25519signed at the client1×exactly-once DAG write0code changes per verticalRead the teardown →
Founder & Principal Engineer · Aug 2025 – PresentAn AI companion that doesn't forget youA companion device has two hard requirements that pull against each other: it has to remember you across months, and it has to answer fast enough to feel like conversation rather than a query. I built the platform end to end — custom ESP32-S3 hardware, the memory system, and a multi-tenant cloud backend licensed to third-party brands as a white-label brain. Memory is three tiers — Redis for anything needed in under 5ms, a vector store for semantic and emotional recall under 50ms, a knowledge graph for structured relationships under 150ms — queried in parallel every turn. Episodic memory decays on a 30/90/180-day schedule by domain, but anything a child returns to repeatedly gets promoted into the permanent knowledge graph, so what actually matters never ages out even though the raw log does. The voice path is streaming end to end, budgeted at 800ms and tuned to hit 520ms. 520mslatency target800mshard budget3memory tiers3xaccesses to promoteESP32-S3custom hardwarePer-childvector + graph isolationRead the teardown →

The through-line

Systems you can operate, not just demo

Is it replayable?Or is every incident archaeology?
Does the state machine have a path backwards?Monotonic state is cacheable state.
Is the scaling signal honest, or merely available?CPU is available. Backlog per worker is honest.

Nine years

How I got here

2017

Biz2Credit

Software Engineer

Monolith to microservices. React and Redux frontends. The ordinary work that teaches you what breaks.

2019

StoreTasker

Software Engineer

A Ruby proxy layer for request capture, and leading the move to React and Node.

2021

Genentech

Software Engineer

ML platform tooling in biotech — a model-federation dashboard and the training topology as an interactive graph.

2023

Remotasks (Scale AI)

NLP & Prompt Engineering

Prompt strategies and evaluation loops for LLM output quality.

2024

YellowPad

Founding Engineer

Legal-AI startup. Platform from zero, 2,000 users, SOC 2 Type 2.

2025

Cere Network

Senior Software Engineer

A ten-agent replayable NLP pipeline on a Substrate-backed decentralized data platform.

2025

Builders Studio

Senior Platform Engineer

A multi-tenant conversation-intelligence platform — nine queue-driven workers, retrieval, and an MCP server.

2025

Sunny Labs

Founder & Principal Engineer

My own company. Custom hardware, three memory tiers, a voice loop tuned to a 520ms target.

Writing

Production teardowns

All 16 →