Portrait of Nardi Xhepi

ML Research Engineer
at Doctolib.

I'm an ML Research Engineer at Doctolib. I hold an M.S. in Data Science from École Polytechnique and an engineering degree from Télécom SudParis.

My main interests are large language models, time series, and computer vision.

I speak French, English, and Albanian.

Experience

Download full CV ↓

Doctolib

ML Research Engineer

April 2026 – Present

  • Phone Assistant (production): shipped a streaming end-of-utterance detection model on continuous audio, raising recall from 56% to 90%.
  • Consultation Assistant (clinical safety): built evaluation pipelines on a production golden dataset, with calibrated LLM-as-judge scoring and statistical drift monitoring.
  • Speech stack: benchmarked and fine-tuned speech-to-text foundation models (NVIDIA NeMo, Parakeet) and evaluated text-to-speech models, weighing accuracy, latency and serving cost; deployed in production with Ray Serve.
  • Data quality: ran annotation campaigns with human review, and diagnosed a pipeline bug that was corrupting 12% of the evaluation data.
Evaluation LLM-as-judge Speech Ray Serve

Dassault Systèmes

Research Scientist Intern

Feb 2025 – Aug 2025 · LLMs & AI Agents, ENOVIA

  • Hybrid retrieval: built RAG pipelines on Qdrant and knowledge graphs, combining BM25 with neural early- and late-interaction retrievers.
  • Agents and extraction: developed LLM agents for entity and relation extraction on enterprise data; benchmarked open-weights models (Mistral Small, Llama 3.1, SmolLM 3B) for the trade-off between accuracy and compute.
LLM RAG Knowledge Graphs Qdrant

EDF R&D

Research Scientist Intern

Jul 2024 – Jan 2025 · Time Series & Anomaly Detection, DOAAT

  • Benchmark to production: designed ensemble detectors for multivariate industrial time series, improving AUC-PR by 15% over baselines, with up to €60M in avoidable annual losses at stake. Published at ICDE 2026.
  • Representation learning: built feature-based latent representations of time series to drive detector aggregation and transfer from benchmarks to production.
Time Series Anomaly Detection Ensemble Methods Python

Publications

From Benchmarks to Production: Transferring Time Series Anomaly Detection Methods for Electricity Production Monitoring

N. Vautier, P. Caron, N. Xhepi, F. Bizeul, M. Boumghar, et al.

IEEE ICDE 2026 · Montréal, Canada · hal-05630277

Education

École Polytechnique

M2 Data Science, highest honors

Sep 2024 – 2026 · Palaiseau, France

Télécom SudParis

Diplôme d'ingénieur, Computer Science & Data Science

Sep 2022 – 2026 · Évry, France

Skills

Production AI

Prototype to production Evaluation harnesses Observability Latency and cost Streaming inference

LLMs & Agents

RAG (hybrid retrieval) AI agents Fine-tuning Knowledge distillation Pseudo-labeling

Evaluation

LLM-as-judge calibration Golden datasets Drift monitoring Human annotation

Speech & Time Series

STT / TTS Anomaly detection Representation learning

Engineering

Python PyTorch Transformers scikit-learn Ray Serve Qdrant Git Bash Linux

Languages

French — Fluent English — Fluent Albanian — Native