Formare AI

Implementation articles

How each project was built

Technical walkthroughs of the architecture, the code, and the decisions. Each article covers one production project from problem to deployment.

June 2026

  • AI AgentsJun 114 min

    Deploying Production-Grade Multi-Agent Systems with Google ADK on Vertex AI

    Up to 74% of AI projects stall between prototype and production. This guide covers the exact architecture patterns we use in enterprise retail environments to run 5+ concurrent AI agents in production — demand forecasting, personalization, and operational intelligence — all on Google Cloud with full audit trails and GDPR compliance.

    • Google ADK
    • Vertex AI
    • Cloud Run
    • A2A Protocol
    • Gemini

May 2026

  • Data EngineeringMay 2011 min

    Optimizing BigQuery and Gemini API for Real-Time Enterprise AI Pipelines

    Unoptimized LLM tool calls against BigQuery can cost €3,000+/month at enterprise scale. This guide covers six specific optimizations used in production in enterprise retail environments — partition strategies, context caching, materialized view pre-computation, and batched embedding generation — that cut costs by 70% while dropping P95 latency below 1 second.

    • BigQuery
    • Gemini API
    • Vertex AI
    • Cloud Run
    • Cost Optimization

  • AI AgentsMay 513 min

    EU AI Act and GDPR Compliance for RAG Systems on Google Cloud

    The EU AI Act entered enforcement in 2026 with fines up to 3% of global annual turnover. GDPR carries an additional 4%. For European companies deploying RAG systems and AI agents, this article covers the exact Google Cloud architecture that satisfies both — data residency, model explainability, audit logging, and the five code-level controls that most teams miss.

    • EU AI Act
    • GDPR
    • RAG
    • Vertex AI
    • Google Cloud

May 2025

  • Full-Stack AIMay 1514 min

    Building a Conflict-Safe AI Booking Platform with Gemini and PostgreSQL

    A UK private driver was managing all bookings via WhatsApp. We replaced that with a Gemini agentic loop with 11 tools — geocoding, fare calculation, conflict-safe booking with PostgreSQL advisory locks, Stripe payments, Google Calendar, automated emails, and real-time flight tracking. Zero double-bookings since launch.

    • Gemini
    • PostgreSQL
    • A2A Protocol
    • MCP
    • Stripe

  • Voice & BrowserMay 18 min

    WebMCP Data Extraction Agent: Replacing Fragile Scrapers with Chrome's Tool Protocol

    Traditional web scraping breaks whenever a website changes its HTML. WebMCP (Chrome 146+) lets websites expose structured tools that AI agents can discover and invoke directly — like a REST API, but built into the browser. We built an extraction agent using CDP + Gemini planning + schema validation.

    • WebMCP
    • Chrome DevTools Protocol
    • Gemini
    • Node.js
    • TypeScript

April 2025

  • Retail AIApr 2011 min

    Multimodal Product Search Under 300ms with Vertex AI Vector Search

    Grocery search fails when customers don't know the product name. We built a multimodal assistant using Vertex AI multimodal embeddings, ANN vector search, and the Ranking API — achieving <300ms P99 latency and >95% Recall@20 across 10,000+ products.

    • Vertex AI
    • Vector Search
    • Gemini Live API
    • Multimodal Embeddings
    • Ranking API

  • AI AgentsApr 127 min

    Building a Romanian Market AI Assistant with Google ADK, Real-Time Search, and MCP Reservation Tools

    General AI assistants give generic answers for local questions because they have no real-time access to local reviews or the ability to complete actions. We built a Romanian market assistant with Google Search grounding for current Bucharest restaurant recommendations and a custom MCP tool for ialoc.ro reservations.

    • Google ADK
    • Gemini 2.5 Flash
    • MCP
    • Google Search
    • FastAPI

  • Voice & BrowserApr 510 min

    Real-Time Voice AI with Gemini Live API: PCM Streaming, React Hooks, and Three Critical Bugs

    Building real-time voice AI with Gemini's native audio streaming looks simple in the docs. In practice, three non-obvious bugs will stop you: PCM alignment from Int16Array views, stale useRef values in audio callbacks, and React 19 ref callback returns. Here's the correct implementation.

    • Gemini Live API
    • WebSocket
    • WebAudio API
    • React 19
    • TypeScript

March 2025

  • AutomationMar 258 min

    Building a LinkedIn Content Agent with MCP Server, OAuth 2.0, and Gemini

    LinkedIn's API is complex, its OAuth flow is non-standard, and managing tokens manually is error-prone. We built a complete automation suite: feed monitor with scoring, AI post generator, and an MCP server that exposes LinkedIn actions to any AI agent.

    • LinkedIn API
    • MCP
    • Google ADK
    • Gemini
    • OAuth 2.0

  • Retail AIMar 109 min

    Building a Multi-Agent Retail Merchandiser with Google ADK and Vertex AI

    Retail merchandising decisions — what to feature, how to rank search results, which products are substitutable — require expertise across thousands of SKUs. We built a multi-agent ADK system that codifies this expertise: Planner → parallel Vector Search + Re-ranker → Feature Store enrichment → Gemini strategy layer.

    • Google ADK
    • Vertex AI
    • Ranking API
    • Feature Store
    • FastAPI

February 2025

  • Voice & BrowserFeb 288 min

    Building a Context-Aware Chrome AI Extension with Manifest V3 and Gemini

    A Chrome extension that detects what kind of text you've selected — code, email, product description, data table, address — and serves Gemini-powered actions specific to that context. Built with Manifest V3 Service Worker, content script context detection, and a shopping variant with competitor price comparison.

    • Chrome Extensions
    • Manifest V3
    • Gemini
    • Service Worker
    • TypeScript

  • Data EngineeringFeb 1810 min

    Building a Self-Hosted BigQuery Data Catalog on GKE Autopilot

    Large data teams on BigQuery have no idea what tables exist or what columns mean. We built a self-hosted data catalog that syncs metadata directly from BigQuery, generates AI column descriptions with Gemini, shows lineage graphs, and deploys on GKE Autopilot with Workload Identity — no stale Confluence docs.

    • BigQuery
    • FastAPI
    • GKE Autopilot
    • Gemini
    • Data Lineage

  • Data EngineeringFeb 59 min

    Transforming BigQuery Product Tables into Knowledge Graphs with Neo4j and Gemini

    Relational databases store product relationships in flat tables. Knowledge graphs make them first-class citizens. We built a pipeline that extracts product data from BigQuery, infers graph edges from taxonomy, embeddings, and basket co-occurrence, loads into Neo4j, and adds a Gemini NL→Cypher agent for natural language graph queries.

    • BigQuery
    • Neo4j
    • Gemini
    • Vertex AI Embeddings
    • NetworkX

January 2025

  • Data EngineeringJan 258 min

    Natural Language BigQuery: Building an NL→SQL Multi-Agent System with Google ADK

    SQL is a barrier for analysts and a bottleneck for data teams. We built a two-agent system using Google ADK: a schema-aware BigQuery agent that translates natural language to SQL (preventing hallucinated column names), chained with a sandboxed Pandas agent for statistical analysis.

    • Google ADK
    • BigQuery
    • NL2SQL
    • Pandas
    • Code Interpreter

  • AutomationJan 109 min

    Automated SEO Blog Pipeline: Gemini + CRAAP Test Source Verification + WordPress REST API

    Publishing consistent, research-backed blog content takes 4–6 hours per article. We automated the entire pipeline: multi-source research with CRAAP test verification, Gemini generation, SEO integration, and WordPress publishing — 3–5 minutes per article, running daily. V9 in production.

    • Gemini
    • Vertex AI
    • WordPress REST API
    • CRAAP Test
    • Cloud Run

December 2024

  • AutomationDec 157 min

    Automated Newsletter Summarizer with Gemini Search Grounding and Sentiment Analysis

    Keeping up with news across multiple technical topics manually is unsustainable. We built an automated pipeline using Gemini with Vertex AI Search grounding that retrieves current articles per topic, generates verified summaries with sentiment signals, and produces newsletter-ready Markdown — running daily on Cloud Scheduler.

    • Gemini
    • Vertex AI
    • Google Search Grounding
    • Cloud Scheduler
    • Cloud Run

November 2024

  • AI AgentsNov 207 min

    Building an AI Storytelling Platform for Children with Gemini, Streamlit, and Panel

    Children's storytelling AI has unique requirements: narrative consistency across chapters, age-appropriate content controls, interactive branching, and character memory. We built MyStoryLand in two UI variants (Streamlit for speed, Panel for interactivity) with Gemini 2.5 Flash handling both story planning and generation.

    • Gemini 2.5 Flash
    • Streamlit
    • Panel
    • Cloud Run
    • Python

October 2024

  • Data EngineeringOct 307 min

    YouTube Comment Intelligence with YouTube Data API v3, Streamlit, and Plotly

    YouTube Studio analytics are surface-level. We built a Streamlit dashboard that exports all comments with pagination, runs sentiment and topic analysis, detects questions for content ideas, and plots engagement decay curves — with full YouTube Data API v3 quota management.

    • YouTube Data API
    • Streamlit
    • Plotly
    • Pandas
    • Python

Request a discovery call

30 minutesGoogle MeetFree