Romain AndreAI Engineer

Context

Product/client recommendation system for sales teams. The engine works in both directions: given a client, which products to recommend? Given a product, which clients to target? Exposed via a REST API with cold-start handling and business rule filtering.

Approach

• Two-stage architecture: Two-Tower Model for vector retrieval followed by a Deep & Cross Network Reranker for final scoring
• Behavioural feature engineering (RFM) combined with product text embeddings, training pipeline with strict temporal split
• Bidirectional REST API with popularity fallback for cold-start and business rule filtering

Results

Tech stack

Context

Design of a retrieval-augmented generation system to enrich an AI assistant's answers with internal company data. The pipeline had to be robust, scalable and serve users in real-time.

Approach

• Full RAG architecture: document ingestion, vector indexing, retrieval and generation
• Reasoning chain orchestration with LangChain and LangGraph
• Exposed via FastAPI + WebSocket API for streaming interactions
• Performance monitoring and evaluation with LangSmith

Results

Tech stack

Context

Brazilian jiu-jitsu decision-making is highly contextual and sequential — intermediate practitioners often struggle to know what to do next in a given situation. As a competitor, I designed RollSenseAI to combine symbolic reasoning (knowledge graph from flowcharts) with LLM power, delivering explainable decision support, grappling-specific conditioning advice, and structured technique explanations.

Approach

• Orchestrated multi-agent architecture: Decision Agent (knowledge graph / flowcharts), Conditioning Agent (RAG + training documents), Learning Agent (books, IBJJF/ADCC rules), Clarification Agent, Safety Agent
• Knowledge graph modeling Brazilian jiu-jitsu positions, states and actions with LEADS_TO, REQUIRES, COUNTERS relationships — enabling deterministic traversal and explainable responses
• RAG pipeline (chunking, embedding, vector search) feeding conditioning and learning agents with authoritative reference sources
• Central orchestrator managing intent classification, inter-agent routing and conversation state (position, gi/no-gi, level)
• Dedicated Safety Agent to detect medical questions, add disclaimers and prevent unsafe training advice

Results

Tech stack

Context

Smart Live Prod is a collaborative project funded by Bpifrance and the Région Bretagne, bringing together Electric Brain, INSA Rennes' VAADER team, and Vodalys. The goal: make live video broadcasts accessible to ~300,000 LSF speakers in France by automatically generating French Sign Language translation from spoken text. I was responsible for Lot 2 — the LSF translation module — in collaboration with INSA Rennes.

Approach

• Modular pipeline text → gloss → pose → video, following the state of the art in automatic sign language generation
• Text2gloss module: NMT Transformer model (OpenNMT) trained on multilingual corpus, with progressive fine-tuning toward LSF and data augmentation (back-translation, self-training)
• Gloss2pose module: structured gloss → pose lexicon, contextual semantic search via sentence embeddings (HuggingFace), smooth transitions between signs
• BLEU/ROUGE evaluation for text2gloss, human evaluation by LSF speakers for gloss2pose

Tech stack

Context

International Kaggle challenge aiming to predict the distribution of ~10,000 plant species across Europe. The goal: leverage heterogeneous environmental data (satellite, climate, soil) to model species presence, with direct applications in biodiversity conservation.

Data

• Sentinel-2 satellite imagery (RGB + NIR, 10m/pixel resolution, ~1M 128×128 patches)
• Landsat multispectral time series spanning 20 years (6 bands, 84 seasons)
• Bioclimatic time series (temperature, precipitation)
• Tabular environmental variables: bioclimatic, soil, elevation, land cover, human footprint
• 90,000 PA (presence-absence) and 5M PO (presence-only) observations

Approach

• XGBoost with Bayesian optimization (HyperOpt/TPE) on tabular environmental data
• Dual-branch ResNet-50 (RGB + NIR) with ImageNet transfer learning for satellite imagery
• ResNet-18 for spectral and bioclimatic time series
• Multimodal fusion: concatenation of 4 branches (tabular MLP + 2×ResNet-18 + ResNet-50) followed by a fully-connected layer
• Optimized thresholding on micro F1 for multi-label classification (optimal threshold: 0.212)

Results

Tech stack