Building production LLM systems at Tiime | MS Electrical Engineering, INSA Lyon
I design and ship agent architectures, retrieval pipelines, and evaluation frameworks for real-world accounting automation.
Agents Langgraph multi-agent workflows, dynamic model routing, lazy skill loading
Retrieval Hybrid search (BM25 + vector), Cohere reranking, Bayesian confidence scoring
Evaluation LLM-as-judge, golden datasets, batch comparison across model families
Infrastructure Prompt caching (Bedrock), SingleFlight embedding dedup, OpenSearch indexing
| Project | What |
|---|---|
| neo-deep-agent-lab | Deep agent SQL lab with Modal serverless deployment |
| neo-neo-sn62 | Bittensor SN62 — decentralized AI subnet mining & validation |
| steering-research | Where, how, and why activation steering works on instruction-tuned LLMs |
| Gptq-Babai-Quantization | Visualization of GPTQ lattice reduction for weight quantization |
| sparse-moe-torch | Sparse Mixture-of-Experts routing visualization |
- Mechanistic interpretability — activation steering, feature circuits, sparse probing
- Quantization — GPTQ, Babai lattice reduction, mixed-precision strategies
- Mixture of Experts — routing policies, load balancing, capacity factors
- Information Retrieval — BM25/vector fusion, learned sparse representations, reciprocal rank fusion
- Reinforcement Learning — policy gradients, RLHF/DPO alignment, reward modeling
Python FastAPI Langgraph PostgreSQL pgvector OpenSearch AWS Bedrock Cohere Celery SQLAlchemy
Bordeaux, France