Achraf Ez

I build data and AI systems from a foundation in production machine learning, time series, and engineering.

• Agentic systems
• Retrieval
• Time series
• Practical AI tools

1. Jul 2025-Present

   Co-Founder · LumiereAI

   • Designed, built and shipped a multi-stage LLM pipeline in production, with every model call site behind one provider-agnostic protocol, a deterministic floor that guarantees delivery when the model pass fails, and a measurement harness that cut end-to-end wall clock 49%, serving users at pipelinevideo.com.
   • Designed, built and shipped a multi-stage agentic RAG pipeline in production, with a provider-agnostic model layer behind one environment switch, a deterministic floor that downgrades ungrounded output to not verified rather than guessing, and criterion-level batching that cut 32 sequential LLM calls to about 7 concurrent ones, serving users at auditmalin.com.
   • Built a multi-layer image-forensics engine combining seven classical and deep-learning signals with an LLM judge under a strict structured-output schema, benchmarked at 80% accuracy with 100% precision on hard negatives, recall improved by about 80% relative through threshold recalibration, and awarded two prizes including best overall at the Pioneers hackathon at Station F, live at fidelityai.dev.
   • Python
   • FastAPI
   • Claude Agent SDK
   • OpenAI and Azure OpenAI
   • Google ADK
   • LiteLLM
   • Docling
   • Qdrant
   • ElevenLabs
   • Next.js
   • Vercel

2. May 2024-Present

   Senior Data Scientist · Decathlon

   • Designed and productionized a modeling approach for heterogeneous SKU and store demand series, enabling cross-series learning across about 280k SKUs and 2.4k stores, 69M listed SKU and store pairs and more than 2B rows of history, forecast weekly at an 8-week horizon.
   • Spearheaded SHAP-based explainability for production demand forecasting, building reusable global and SKU-store diagnostic workflows, mentoring teammates, and using feature attribution to uncover model behavior and upstream data-quality issues.
   • Redesigned forecasting evaluation beyond WAPE and aggregate bias with naive-baseline signal, scaled MASE and RMSSE metrics, relative error, and directional SignedBias across dbt and Spark monitoring pipelines, so over-forecasting, under-forecasting, volume, and turnover impact can be judged separately.
   • Led end-to-end feature engineering over weather, price, sales-derived, store and item attributes, from source-data analysis and feature design through ELT pipelines to model integration, backtesting, monitoring, and stakeholder communication, across multiple markets, product universes, and forecast cutoffs.
   • Pioneered internal evaluation of tabular and time-series foundation models, extending the experimentation platform for TabPFN and TabICL and assessing accuracy, interpretability, data requirements, and GPU scalability constraints.
   • Built an agent-ready data-science monorepo using reusable agent skills, MCP connectors, and operational documentation, enabling agent-assisted end-to-end experimentation and a reported 2x acceleration in research and development throughput.
   • Designed a RAG system over multilingual product descriptions: an LLM returns a fixed list of product attributes, retrieval finds the past products a new item most resembles and the product it replaces, and the same text embeddings become features a tabular foundation model can learn from, which is what a launch forecast has to lean on when the item has no sales history of its own.
   • Designed an assistant for forecast support that reads an incoming ticket, gathers the forecast, sales, stock and metric evidence through read-only tools, finds similar cases already resolved, and drafts an answer in which every figure names the query it came from, for a human to approve before it is sent, removing about 7 hours of manual engineering support a week.
   • Python
   • SQL
   • dbt
   • Spark
   • Polars
   • LightGBM
   • MLflow
   • Databricks (Unity Catalog, Asset Bundles)
   • Airflow and Astro
   • AWS
   • Docker
   • SHAP
   • Marimo
   • TabPFN and TabICL

3. Apr 2022-May 2024

   Senior Data Scientist · COMPREDICT

   • Created a novel time series model, FFD, that reduced the company's previous state-of-the-art model size by about 50 times while keeping more than 97% of its predictive accuracy.
   • Built and maintained the internal libraries that kept models deployable to production.
   • Reduced and optimized model sizes for in-vehicle deployment through distillation, quantization, and pruning.
   • Python
   • TensorFlow
   • Keras
   • ONNX
   • Ray
   • Optuna
   • scikit-learn
   • MLflow
   • Kubeflow

4. Jul 2018-Aug 2022

   Data Scientist · Alstom

   • Developed and deployed the company's first deep learning models for time series, predicting train responses to environmental and architectural changes in the railway industry.
   • Led data gathering and custom preprocessing across a team, cutting simulation time and reliance on expensive multibody simulations by 60%.
   • Deployed machine learning models that estimate the cost of tender tasks from historical data and return of experience, which cut the time each tender took.
   • Deployed transformer-based NLP models that unified disparate component classifications, standardizing more than 2,500 parts for downstream predictive maintenance.

5. Jul 2018-Jul 2019

   Data Engineer · Alstom

   • Led statistical and machine learning analyses that informed key design decisions from a railway-dynamics point of view, acting as a point of contact for high-profile clients.
   • Built and deployed a web app for designing suspension components from theoretical and empirical data, using machine learning and signal processing.
   • Implemented a model predicting successful wheel and axle mounting operations, raising the success rate by 20%.

6. Jun 2017-Jun 2018

   R&D Engineer · VibraTec

   • Implemented a signal-processing package centered on convolution techniques.
   • Developed an app predicting wheel and rail forces and estimating the impact of wear and rolling contact fatigue in the railway industry.
   • Co-wrote a paper with LAMCOS at INSA Lyon.

7. Oct 2015-Oct 2016

   Technical Project Manager · Mines Rabat Solar Team

   • Modelled, designed and built a solar-powered car with the student team.
   • Directed the mechanical team through the structural design of the car.
   • Won the Moroccan Solar Race Challenge and took part in COP22.

• Pipeline Video

  Founder and engineer

  Turns a company's website into a finished marketing video: crawl, script, human approval, recorded voiceover, motion graphics, deterministic render.

• LumiereAI

  Co-Founder

  Autonomous agents for enterprise operations.

• auditMalin

  Creator

  AI-assisted Qualiopi pre-audits across the RNQ's 32 indicators, with cited verdicts and prioritized action plans.

• FidelityAI

  Co-creator · team project

  Explainable multi-agent insurance claim triage across documents, photos, voice statements, and policy evidence.