Jimi Vaubien

YourGPT Platform

GenAI platform deployed by Terraform into each client's own Azure subscription, with private networking between every service inside it, modular microservices, and config-driven SPA assembly.

• ▸4 production clients (10-500 users each)
• ▸One deployment per client, in the client's Azure subscription, private networking inside
• ▸Single Terraform codebase supporting arbitrary client Azure configurations

Financial Audit Auto-Tagging

AI agent monitoring employee emails via Microsoft Graph API automating document extraction and audit software integration with Outlook add-in for human-in-the-loop review.

• ▸2 hours saved per agent per day (25% of workday returned)
• ▸Continuous processing of incoming email flow
• ▸Handles plain text, PDF, images, scanned docs

Trade Settlement Agent

Email-based workflow automation with stateful multi-turn conversation management, auto-reinitiation for delayed responses, and intelligent escalation to human operators.

• ▸10 hours/week back-office savings
• ▸Hundreds of counterparty interactions processed daily
• ▸Conservative escalation boundary preventing financial loss

Advanced RAG Systems

Multi-modal retrieval combining hybrid search (BM25 + vector + reranking) with SQL agent and multi-turn agentic coordination via LLM reasoning.

• ▸90% recall with agentic mode vs 30% SQL-only baseline
• ▸~10s average latency for agentic mode
• ▸100s GB of documents at 2 production clients, each with dozens to hundreds of employees

Customer Support LLM Service

Hybrid support architecture on 2x GPU nodes for high availability, with Llama 3 8B (AWQ 4-bit quantization) and vLLM serving. Intent classification with regex parsing (0.1ms) routes 68% of queries to autonomous resolution, 32% to human agents with full context.

• ▸20 QPS peak across 2 GPU nodes (80% peak GPU utilization)
• ▸p99 latency 1.8s end-to-end
• ▸68% autonomous query resolution, zero human touch
• ▸Response time reduced from 1 hour → 45 seconds
• ▸3+ hours saved per human agent per day
• ▸99.9% uptime over 4 months

AI Safety & Guardrails Framework

Defense-in-depth safety system for regulated financial environment: prompt injection classifier (fine-tuned DeBERTa), PII detection for Swiss banking entities, and a red teaming pipeline with 100+ attack variants per deployment.

• ▸99.2% prompt injection detection rate
• ▸0 PII leakage incidents in 6 months production
• ▸2.1% false positive rate on legitimate queries
• ▸150ms average latency overhead
• ▸Passed internal audit, enabling production deployment

Data Platform Migration

Transformed manual SQL pipelines into automated DBT/Dagster architecture with semantic versioning, data quality testing, and unified orchestration of SQL and Python ML nodes.

• ▸Eliminated manual pipeline failures and data corruption
• ▸10s to 100s GB pipelines migrated to fully automated scheduled execution
• ▸CI/CD with semantic versioning enabling safe deployments

Unsupervised Fraud Detection (Banking)

Full anomaly detection system from POC to production with explainable PDF reports: pixel-level context plots, rule-based DSL, isolation forest, and black-box ML ensemble.

• ▸30% precision (30x lift over random baseline)
• ▸1000s of transactions processed daily
• ▸2+ years in production
• ▸Passed external security audits

B2B Recommender System

Two-tower neural network empowering B2B sales reps: query tower (customer/context) + item tower (product embeddings) with offline batch engine producing ranked lists per client.

• ▸+15% AOV uplift
• ▸10K+ products per client
• ▸Passed external security audits

Real-Time Sound Event Detection

Edge and cloud AI for gunshot/accident detection: MobileNet INT8 quantized via TensorFlow Lite on Raspberry Pi (offline edge) + ResNet cloud version with shared LSTM head.

• ▸Client won a Swiss startup competition with this PoC

B2C Recommender System

Neural matrix factorization with three embedding inputs (user, item, temporal context) replacing broadcast notifications with selective personalized delivery.

• ▸4x CTR uplift over broadcast
• ▸100K+ users

Pharma Data Engineering

AWS data lake with Spark, DBT, and NLP layer (fine-tuned BERT for PII removal, NER, sentiment analysis) for Real World Evidence analysis supporting regulatory submissions. Warehouse design across 100s GB to TB datasets.

• ▸100s GB to TB across adverse events, EHR, claims data
• ▸PII removal, NER, sentiment analysis pipelines

Cyclist Detection (Sports)

MaskRCNN instance segmentation for automated finish-line detection in dense peloton scenarios, integrated with photo finish hardware.

• ▸80% accuracy on standard scenarios
• ▸55% on hard scenarios (where humans also struggle)
• ▸Architecture later adapted for gymnastics athlete detection

Job Seeker Platform

Full-stack recommendation system with A/B testing framework deployed on GCP, built for academic research on job seeker behavior.

• ▸Adopted by official Swiss state administrations
• ▸Cohort-based A/B testing framework built from scratch

Unsupervised Defect Detection

One-Class SVM, Autoencoders, and GANs for defect recognition with no positive training examples available.

• ▸6/6 Master thesis grade
• ▸Supervised by Eric Cosatto (NEC Labs)

Geological Simulation Acceleration

Conditional GANs with chunk-based positional conditioning replacing slow physics-based simulation for soil composition generation.

• ▸10x speedup vs physics-based simulation
• ▸Predictions integrated into downstream geological modules

Sea Level Time Series Prediction

RNN/LSTM for reconstructing historical sea levels from soil composition measurements at varying depths.

Embeddable Real-Time Search Widget

React widget included via script tag with recommendation-powered live search ranking.

• ▸Deployed to 10+ production eCommerce clients
• ▸Company became MassChallenge finalist and was acquired