AI/ML Engineer and Cloud Architect with 5+ years engineering production-grade intelligent systems for international clients across healthcare, fintech, and e-commerce. Fluent in the full AI stack — from Speech AI and LLM fine-tuning to GPU cloud deployment and MLOps — delivering measurable business outcomes from Antananarivo to the world.
From GPU-powered speech models to autonomous multi-agent systems, I architect and ship AI that generates real business value.
End-to-end speech pipelines combining ASR fine-tuning, neural machine translation, and voice cloning. Production systems processing 6 languages with sub-2s latency on GPU cloud.
RAG architectures (vector, graph, hybrid, agentic), multi-agent orchestration with LangGraph and Google ADK, and RLHF-optimized pipelines that automate complex enterprise workflows.
Scalable cloud-native platforms on GCP and AWS. Vertex AI Pipelines, Cloud Run GPU, BigQuery data warehouses, and full CI/CD MLOps lifecycles — budget-optimized and production-hardened.
SaaS platforms built from scratch with Next.js, NestJS, WebRTC real-time features, payment integrations, and AI-powered onboarding — shipped with 750+ tests and full DevOps pipelines.
From USAID health data systems to GPU-powered speech AI and UK fintech — building consequential software across three continents.
Selected case studies demonstrating measurable impact across AI, cloud infrastructure, and full-stack engineering.
Engineered a first-of-its-kind linguistic bridge converting spoken French, English, German, Spanish, Italian, and Portuguese into 6 Malagasy dialect voice outputs in real time. The 4-stage GPU pipeline — VAD → ASR → Neural MT → TTS — runs on Cloud Run L4 at just $42/month, making advanced speech AI financially viable for an emerging-market public institution.
Architected from scratch a multi-country telehealth SaaS connecting patients with 17 healthcare provider categories. AI-powered document OCR automates provider onboarding; real-time WebRTC powers video consultations; a configurable workflow engine handles 33+ consultation types with automated role-based notifications. Mobile money payments built-in.
Delivered an AI platform replacing manual pension administration workflows for UK financial advisers. Gemini Flash OCR achieves >95% first-pass accuracy on multi-format pension documents; AI-driven phone automation handles IVR navigation; the compliance engine enforces FCA COBS 9/19 with immutable audit trails and automated suitability reports.
Leading two production AI systems: an autonomous B2B sales agent (Google ADK + LangGraph) that runs the full pipeline from cold outreach to deal closing with RLHF optimization; and a multimodal search platform powered by Gemini 2 Embedding — benchmarked against CLIP — where users search with images or text and pay directly with mobile money.
LangGraph and RAG-powered text-to-SQL solution retrieving SQL patterns, DDL, and metadata for multi-query execution over BigQuery. Fine-tuned SQLCoder-7B-2 on BigQuery dialect.
Master's Thesis: hybrid RAG system combining Neo4j knowledge graph, Pinecone vector search, and Elasticsearch. Multi-modal document and image extraction from cloud infrastructure.
Google ADK and LLM orchestration with multi-source intelligence gathering from social media and company data. Pattern-based investment signal analysis for B2B prospecting.
VLM-based multi-format attachment analysis (PDF, images, tables) for a Swiss industrial client. Microsoft Graph webhooks, SharePoint Excel integration, and LLM-based routing and triage.
NER-based skills extraction with spaCy, NLP semantic matching for candidate-job pairing. Python FastAPI and React.js with advanced recruiter interface with candidate ranking.
YOLO V8 fine-tuned for real-time multi-class retail inventory detection with instance segmentation and TensorRT optimization. REST API with real-time inference pipeline. Presented at AI Event 2024.
Google Dataflow and Apache Beam distributed ETL pipeline vectorizing 105,000 product images into BigQuery. K-means clustering and cross-modal vector similarity search across images and text.
PostgreSQL replicated via Airbyte CDC to BigQuery; dbt SQL transformations; Looker Studio real-time dashboards for mobile biometric kit performance monitoring and BigQuery ML ARIMA+ forecasting.
LangGraph with memory persistence, Vertex AI and BigQuery integration, SQLCoder fine-tuning on BigQuery dialect, and automatic visualization dashboard generation for business intelligence.
Architectural study of ASR and TTS systems for Malagasy as a low-resource language using Chirp (Vertex AI Studio). Established the transformer adaptation methodology later applied in production at Maison du Numerique.
Deployed on GKE with WhatsApp Business API integration. NLP for flight status, navigation, and airport services. BigQuery for query analysis, real-time translation, and multilingual NLP.
No-code configurable chatbot with image recognition on GCP. RAG-based chatbot with configurable modes (image, text, combined). Google Sheets catalog integration with under 3s response time.
DHIS2 platform module automating detection of duplicates, missing values, outliers, and non-applicable entries in national COVID-19 health data using Isolation Forest anomaly detection.
Vaccine stock management and healthcare facility navigation tool for the USAID PMM Program. Interactive maps with turn-by-turn directions, vaccine inventory visualization by type and date.
NLP pipeline for pathology prediction from clinical symptoms across 10 disease classes. Achieved 94% accuracy with XGBoost + CatBoost ensemble, TF-IDF, Word2Vec, and spaCy preprocessing.
Two models in the Pan-African data science competition: a content recommendation model for 122M social media interactions; and a credit default risk classifier trained on real bank lending data using XGBoost and LightGBM.
Evaluated SQLCoder-7B-2, GPT-3.5/4, and Claude on architecture, hallucination mitigation, BigQuery syntax, and Spider benchmark. Produced multi-cloud deployment guide for enterprise adoption.
Whether you're a CEO evaluating an AI strategy, a CTO scaling a platform, or a recruiter building a world-class team — let's talk. I deliver production-grade AI, not prototypes.
atr.guillaume@gmail.com