Four systems, each described with its constraints, architecture and measured outcome.
Replaced a legacy reporter with a queue that holds every event until the parent group acknowledges it, sends it exactly once, and scales from an afternoon trickle to a full day's backlog within minutes.
15M+outbound events a day, plus up to 3M inbound
Google Cloud Spanner Cloud Functions Pub/Sub Cloud Storage
A customer agent on a delivery network's app and website that answers from the live parcel record, investigates "delivered but not received" the way operations does, and passes the right conversations to a person.
10–15kconversations a day
Python LangGraph Gemini Vertex AI Cloud Functions
An assessment service that takes a delivery network's loss claims, lets a model answer three narrow questions, runs every check in a fixed order, and either approves or hands a person the full list of reasons.
200claims a day assessed
Python FastAPI LangGraph Gemini Vertex AI
A proof of concept for a 10,000-person enterprise, run with its 300-person IT department: every model in one chat, a personal document library, the company's policies answerable, and two agent builders, one visual for engineers and one drag-and-drop for everyone else.
300people in the proof of concept, of an enterprise of 10,000+
React Node Python LangGraph MCP
Exactly-once delivery under hard deadlines, queues that scale from an afternoon trickle to a full day's backlog in minutes, and every event accounted for: counted in, counted out, each attempt on record. We have run this at fifteen million events a day.
Google Cloud, Spanner, Pub/Sub, Cloud Functions, Dataflow, BigQuery, Cloud Storage
Customer and back-office agents that read live data, follow the business's own rules, and know when to hand over to a person. Every turn traced, and every model change scored against real conversations before a customer sees it.
Python, LangGraph, Gemini, Vertex AI, MCP, LangSmith, FastAPI, Cloud Run
Workflows where the model answers narrow questions and deterministic rules make the decision, with every check logged for the reviewer. Low confidence is a reason for review, not a guess. Built for claims, applicable wherever a decision has to be explained.
LangGraph, Cloud Tasks, Cloud SQL, BigQuery, Pydantic
The ground the systems above stand on: cloud migrations, backends in Java and Node.js, front ends in React, a platform maintained for 8,000 employees, and a proof of concept delivered with a team of twenty. Ten years of it before the word "agent" meant anything.
Java, Node, TypeScript, React, Google Cloud, PostgreSQL, microservices
A replay of one production request, step by step, as our tracing records it.
Founder and principal engineer
Over ten years building and running software that businesses depend on every day, from developer through solutions architect and engineering manager to principal AI engineer. He has designed data pipelines that move millions of records a day, led teams through cloud migrations, and put LLM agents in front of customers and back-office staff with the evaluation and tracing that production demands. He builds the system, runs it in production and hands it over working.
Experience
2 weeks, fixed price
We read your data, your incidents and your constraints, then write the architecture, the evaluation plan and a build estimate you can take to another vendor. If a smaller system would do, the estimate says so.
weekly demos, measured
A thin end-to-end path first, then depth. Every week you see the system on your data with the metric that matters next to it. Nothing is "done" without a number.
your team or ours, retainer in hours
It stays with your team, or we run it for you. Either way we write the runbooks, rehearse the failure modes and stay available on a retainer sized in hours a month. The code, the infrastructure and the evaluation set are yours from the first commit.
Systems run in your own cloud account and region, typically Google Cloud in the EU or the UK, under agreements that satisfy GDPR; your data never leaves that account. Model providers are used only through endpoints that do not train on submitted data. We document data flows, retention and human oversight for each system.
Write to Jakubai@byteandreason.plWe reply within one working day.
