Google Cloud AI engineering
Formare AI takes Gemini and Google ADK agents from prototype to a production system: persistent sessions, tool reliability, tracing, cost control, EU data residency and a handover you own. Led by the engineer who writes the code. Fixed-scope projects from €8,000, delivered in 2–12 weeks.
Bucharest, RomaniaWorking with clients worldwide
Data
Pipelines, warehouses, features
Intelligence
Models, agents, retrieval
Services
APIs, auth, payments, jobs
Interfaces
Web apps, dashboards, chat
How we work
Three commitments apply to every engagement, whether it starts as a review of an agent you already have or as a new build.
Agents fail on state, tools, tracing and cost far more often than on the model. Those are built in from the first sprint.
Gemini, Vertex AI, ADK and BigQuery on Cloud Run or GKE, with the region locked at project level before the first row is written.
Day rate €600–900, fixed projects from €8,000, retainers from €4,500 a month. The engineer who scopes the work writes the code.
Teams with an agent prototype
A notebook, demo or pilot works, but it is not something you can run, secure and support.
Where we start: An agent production review: a written, ranked plan, then a hardening project with persistent sessions, tracing, access control and a deployment you own.
Startups and product teams
You need to launch an AI product and have no in-house ML or cloud engineering team.
Where we start: A fixed-scope MVP: the core workflow running end to end on Cloud Run is the first milestone.
Data and platform teams on Google Cloud
You already run Gemini, Vertex AI or BigQuery and need senior help on architecture, cost or compliance.
Where we start: The 4-week architecture audit, or a BigQuery AI project, then a retainer.
Selected work
6 builds: 1 client project, 1 hackathon entry, 3 reference builds and 1 prototype. Each one links to its source code or live site, says which kind it is, and names the service you would hire us for.
Full-StackClient project
George Travel — georgetravel.co.uk
Full-stack AI booking platform for a UK private driver. Gemini agentic loop with 11 custom tools handles geocoding, fare calculation, conflict-safe booking (PostgreSQL advisory locks), Stripe payments, Google Calendar, automated emails, real-time flight tracking, and weather. Interoperable via A2A protocol and MCP.
Production figure from the live system, reported from launch to the date of the build article (15 May 2025). Conflicts are prevented by a PostgreSQL advisory lock around the availability check and the insert.
Hire us for this: AI agent development on Google Cloud. An agent that quotes, books and takes payment for your service business, against your own calendar and rules.
AI AgentsReference build
Gemini Act — Google ADK agent inside Google Workspace
A Google ADK agent that takes real actions from Google Chat. People message it or @-mention it in a space, and it works across their Gmail, Calendar, Drive and Chat through Workspace MCP servers, plus your own internal APIs as tools. Each person signs in with their own Google account (OAuth refresh tokens in Firestore), so one shared agent serves everyone without sharing sessions. The webhook checks the bearer token Google signs, acknowledges at once and posts the answer through the Chat API afterwards, because an agent loop outlasts Chat's roughly 30-second reply window. The Workspace MCP servers it relies on are in Google's Developer Preview, not generally available.
Hire us for this: AI agent development on Google Cloud. We deploy it into your Workspace domain and connect your own systems as tools, with confirmation before any write, send or delete.
DataPrototype
AI Data Studio — Gemini and Google ADK over BigQuery
Turns a plain-language request into a validated dashboard spec, maps it onto a Looker Studio template and returns a link that opens the report wired to a BigQuery table. Gemini plans only against a semantic model of the table, so it cannot invent fields; a deterministic validator, a BigQuery allowlist and sensitive-field blocking sit between the model and your data. A conversational ADK analyst agent runs read-only SQL that is dry-run first, SELECT-only and capped at 1 GiB billed by default. Exercised against an energy-market dataset; not yet deployed to Cloud Run.
From the repository: its README reports 21 passing tests, and the cost cap is the default in its configuration. Not production figures.
Hire us for this: BigQuery data engineering with AI. We take it to production on your tables: semantic models, dashboard templates, access control and a Cloud Run deployment.
DataReference build
Self-hosted BigQuery data platform
Self-hosted data catalog for GCP / BigQuery teams. FastAPI + React + PostgreSQL + GKE Autopilot. AI-powered column descriptions via Gemini, data lineage via Cloud Data Lineage API, quality scoring, OAuth2, and full table/column search. No stale Confluence docs — metadata syncs directly from BigQuery.
Hire us for this: BigQuery data engineering with AI. We deploy it into your GCP project, connect your BigQuery sources and adapt it to your organisation. The repository documents how.
AI AgentsReference build
Factory Business AI agent — Google ADK and MCP
A Google ADK business-intelligence agent for entrepreneurs, investors and consultants. It calls one MCP server with 26 tools: Romanian company-registry data (profiles, financials, administrators, competitor comparison, risk assessment) from the Targetare API, and Google Maps tools for competitor density, accessibility, distances and place details. Google Search covers market research. A React chat front end shows each tool call as it streams, and the backend ships with Cloud Run and GKE deployment scripts using Workload Identity and Identity-Aware Proxy. The company data needs the client's own Targetare API key.
Hire us for this: AI agent development on Google Cloud. We build the same pattern on your data sources: your registry, ERP or internal APIs behind an MCP server, with the agent and chat interface in front.
Retail AIHackathon entry
Gemini Live Agent Challenge — Hackathon
Grocery product search on a catalogue of about 10,000 products, built for the Gemini Live Agent Challenge. Image + text search on Vertex AI embeddings and Vector Search, in three variants: text, image, and live audio over WebSocket with the Gemini Native Audio API.
Not production measurements. The latency and recall figures are targets for this index configuration, taken from Google's published Vector Search benchmarks. The precision gain is the figure Google reports for the Ranking API.
Hire us for this: Retail and e-commerce AI on Google Cloud. We build the same pipeline on your catalogue: embeddings, a Vector Search index, Ranking API re-ranking and an API in front of it.
Process
A predictable process for founders and engineering leads who need it done right the first time.
30 min, free
Tell us what you're building. We'll tell you if it's feasible, what stack fits, and what it will cost. We'll also tell you if GCP is overkill for your use case — most consultants won't.
Within 48 hours
Clear scope, milestone breakdown, and a fixed price or day-rate. No vague estimates, no surprise invoices. You know exactly what you're getting.
First milestone
Before the full build, we ship the smallest version that proves the idea: the core workflow working end to end on Cloud Run. You test it, we see what holds up, and you decide what goes into the full build.
2–12 weeks
Working code from week one, not a demo on the last day. Cloud Run deployments happen in the first sprint so you're testing against real infrastructure, not a localhost mock.
30-day support included
Full documentation, runbooks, cloud credentials, and 30 days of email support after delivery. You own everything: code, infrastructure, and knowledge.
Engagement
Transparent pricing with no retainer lock-in. The discovery call is always free.
Entry offer
Your agent works in the demo. We read the code and the deployment, test the failure modes that appear in production (state, tool reliability, tracing, cost, security, EU data handling) and give you a written, ranked plan with effort estimates. It is read-only: nothing in your systems changes.
Fixed scope, quoted in writing
Written planeffort estimatesproposal within 48 hours
Architecture audit
Not sure if you need agents, a pipeline, or just a cleaner data model? We audit your current setup, map what's achievable on your stack, and deliver a written architecture plan, prioritised by ROI rather than technical interest. No vendor bias. Most clients move to a build engagement after this.
From €8,0004 weeksfixed scopewritten deliverable
All rates in EURVAT may apply
Technology
64 technologies in 6 groups, from Gemini and ADK at the AI layer to PostgreSQL and Docker underneath. Each one links to the write-ups that document it.
The full stack and the builds behind each technology
All production deployments on Google Cloud Platformeurope-west1/2/4Romania & UK
Our approach
A principal-led AI engineering practice on Google Cloud, based in Romania and working worldwide.
We take on a limited number of engagements at a time: the ones where the problem is genuinely hard and the client values production quality over speed.
Formare AI is a Google Cloud AI engineering practice in Bucharest, Romania, led by Petrică Rădan. It designs, builds and deploys AI agents and data platforms on Gemini, Vertex AI, Google ADK and BigQuery, with EU-region data residency, audit trails and a documented handover. The controls it builds in are designed to support GDPR and EU AI Act obligations; whether a system complies depends on how it is used and classified. Formare AI shows 6 selected builds (1 client project, 1 hackathon entry, 3 reference builds and 1 prototype), each linked to its source code or live site and labelled so you can tell client work from demonstrations.
We work primarily in the Google Cloud and Gemini ecosystem and build frontends in React and TypeScript. Python for backends and data pipelines, Node.js and Express for API layers when the project calls for it.
Retail & e-commerce
Multimodal product search and re-ranking
Transportation
AI booking, payment and scheduling agents
Energy
Dashboards and analytics on market data
Business research
Company-registry and location intelligence agents
Articles
Guides and walkthroughs with the architecture decisions, the code and the trade-offs. Each is labelled as a client project, reference build, hackathon entry, prototype or reference guide.
Reference guide14 min
What breaks when a multi-agent system leaves the notebook, and the session, tool-reliability, tracing and cost controls that fix it, with example code for Cloud Run, Cloud SQL and BigQuery.
Reference guide8 min
Thirty checks across state, tool reliability, observability, cost, security and EU data residency, each with why it matters and how to verify it.
Client project14 min
A UK private driver managed all bookings over WhatsApp. A Gemini agentic loop with 11 tools now handles geocoding, fares, conflict-safe booking with PostgreSQL advisory locks, Stripe payments, Calendar and flight tracking. No double-bookings from launch to the article date (May 2025).
Contact
Three quick questions, then your contact details. We'll come back with an honest assessment, not a sales pitch.
30 minutes on Google Meet, free. Honest answers on feasibility, stack, and cost.