Yanis AmirouFrom Applied Science to Production with High ROI
PhD-level expertise with a proven track record securing critical AI initiatives for market leaders
Specializing in GenAI Agents & Forecasting
Track Record
Turning Science into Value
The three-pillar methodology that transforms AI research into production systems generating millions in value
Applied Science
From Theory to Insight
Translating cutting-edge research into practical, explainable models grounded in mathematical rigor.
- Robust uncertainty quantification (Conformal Prediction for risk control)
- Domain-driven architectures using graph-based reasoning
- Custom modeling beyond AutoML for complex constraints
Industrial Engineering
From Prototypes to Scalable Systems
Bridging advanced models and real-world reliability through production-grade MLOps and architecture.
- GenAI systems: Agents, RAG, MCP, data security & GDPR compliance
- Forecast intelligence: Probabilistic predictions, interpretability, and drift monitoring
- Cloud-native deployment: Kubernetes, CI/CD, and cost-efficient scaling
Strategic Impact
From Systems to Measurable Value
Turning technological sophistication into tangible ROI and operational excellence.
- €35M annual savings through supply chain optimization
- Drastic process acceleration: weeks to minutes
- Data-driven decision tools for high-stakes executive contexts
Engagement Solutions
Three flexible engagement models designed to fit your AI maturity and business needs
Strategic Audit
De-risk your AI roadmap
Technical feasibility studies, architecture audits, and strategic guidance for AI initiatives. Perfect for validating approaches before major investments.
What you get
- AI readiness assessment
- Architecture review & recommendations
- Technology stack evaluation
- ROI estimation & risk analysis
Build Your AI
Rapid prototyping to production
End-to-end development of GenAI agents, forecasting systems, or custom ML models. From research paper to working prototype in weeks, not months.
What you get
- Proof-of-concept development
- Custom model training & fine-tuning
- Integration with existing systems
- Documentation & knowledge transfer
Production Scaling
Scale & industrialize AI systems
MLOps infrastructure, CI/CD pipelines, monitoring, and production deployment. Transform research prototypes into reliable, scalable systems.
What you get
- MLOps pipeline setup (CI/CD, monitoring)
- Cloud infrastructure (GCP, AWS)
- Model retraining & drift detection
- Team upskilling & best practices
Or, let's define the optimal roadmap for your specific context.
Schedule a Free ConsultationTestimonials
Validated by industry leaders in AI Architecture and Applied Science
"Yanis has played a pivotal role in building our country-level demand forecasting system and has made outstanding contributions to our deep learning forecasting models, particularly the Temporal Fusion Transformer (TFT) model."
"He built an end-to-end AI agent solution that I was able to safely validate for production. Being able to rely on an expert like Yanis is invaluable: his autonomy and technical rigor greatly facilitate the process of architectural validation."
Want to see similar results for your team?
Let's Discuss Your ProjectSelected Impact Stories
Transforming cutting-edge AI research into measurable business outcomes
Senior AI Engineer - Independent Consultant
Agent "Onboarding Wholesale" (End-to-End)
(POC) Multi-Agent System for IT Support
Senior Data Scientist - Forecasting & Supply Chain - Independent Consultant
Foundation Models Approach
Fine-tuning of SOTA foundation models (TimesFM, Chronos)
Multimodal Forecasting
Designed semantic enrichment pipeline. Generated text embeddings (product descriptions) injected as covariates to capture similarities between items beyond simple time series
Probabilistic Forecast & Reliability
Implemented Conformal Prediction to generate statistically robust confidence intervals. This rigorous uncertainty quantification enables optimal safety stock sizing where classical models fail
R&D & Architecture — Critical to €35M Savings
Refactored TFT architecture with interpretability modules (Feature Importance, Attention Weights) and flexible RNN configurations. These modifications were essential to achieving €35M annual savings: interpretability enabled supply chain teams to trust AI recommendations at scale, while architectural flexibility improved accuracy across diverse product categories
Open-Source Contributions
Integration of these improvements directly into reference library Nixtla/NeuralForecast (PR #1230 & #1104 merged), validating approach robustness
Measurable Impact
Senior Data Scientist | MLOps
Industrialized photovoltaic plant production prediction to optimize energy management and anomaly detection. Main challenge: structuring abundant R&D into robust MVP product, capable of handling complex time series (weather and energy data) and scaling on cloud.
Leadership
Led MVP roadmap (Jira) and mentored junior Data Scientists team to structure delivery and strengthen best practices
Modeling
Trained and optimized boosting models (XGBoost, LightGBM) and Deep Learning for complex time series, including Bayesian Optimization via SageMaker and rigorous performance validation
End-to-End MLOps
Set up complete CI/CD pipeline (GitHub Actions, Docker images, Airflow orchestration) and deployed observability system for data drift and concept drift detection
Successful production deployment of automated forecasting pipeline. Delivery reliability through model quality monitoring (Drift detection) and team upskilling on MLOps standards
Data Scientist
Developed AI solutions for connected insoles analyzing gait and running to detect pathologies and optimize performance. Data from highly noisy sensors, limiting signal precision.
Deep Learning & Signal Processing
Developed deep learning models for IMU signal denoising
Pathology Detection
Feature extraction and gait pathology detection (XGBoost, LightGBM), validated by statistical tests. Scientific collaboration (Physicians)
GenAI & Data Augmentation
R&D Assistant (RAG): Internal chatbot (Llama 2, LangChain, ChromaDB) to query technical documentation. Synthetic Data: Used GANs and Reinforcement Learning (PPO in MuJoCo) to address data scarcity
Research
Implemented cutting-edge architectures (LSTMs, Transformers) for biomedical time series forecasting
Significant noise reduction in signals which improved all machine learning solutions performance. Created dashboards (Streamlit/Power BI) for clinical KPIs monitoring. Technical mentoring of Data Science interns.
Data Scientist R&D
Spatio-temporal forecasting on interconnected sensor networks. Challenge: surpass classical models by exploiting underlying graph structure.
Achieved State-of-the-Art performance on time series forecasting on graphs and validated new modeling approach for energy networks
Solution: • SOTA Hybrid Architecture combining GCN (spatial) + Transformers (temporal) • Graph Convolutional Networks to capture sensor network topology • Transformer layers for temporal sequence modeling • Topological Data Analysis (TDA) for robust structural feature extraction
Industry Recognition & Thought Leadership
My code powers forecasting for thousands of data scientists via Nixtla's NeuralForecast
2.8K+
GitHub Stars
2
Merged PRs
Nixtla/NeuralForecast
State-of-the-art time series forecasting library with deep learning models
Motivation & Context
These TFT modifications were critical to achieving ~€35M in annual savings at Decathlon. The standard implementation lacked two key capabilities: (1) Interpretability — supply chain stakeholders refused to act on 'black box' predictions, so I built native feature importance and attention weight extraction to make recommendations auditable; (2) Architectural flexibility — diverse product categories required different modeling approaches, so I added GRU support, multi-layer stacking, and static covariate initialization. After validating these improvements on real industrial data (where they directly enabled stakeholder adoption at scale), I contributed them back to Nixtla's NeuralForecast library. These enhancements are now used by data scientists worldwide for production forecasting.
Merged Pull Requests
Custom RNN Layers for TFT - Enhanced Modeling Power
Description:
Extended the Temporal Fusion Transformer architecture to handle complex forecasting patterns with three major enhancements: (1) Flexible Multi-Recurrent Layers - Added support for GRU (lighter and faster) alongside LSTM, giving practitioners more options based on their computational constraints. (2) Deep Modeling - Enabled stacking multiple RNN layers to capture higher-order temporal dynamics in complex time series. (3) Custom State Initialization - Implemented state initialization from static covariates, significantly boosting performance on new or sparse series by leveraging cross-sectional information.
Impact:
Enables more flexible architecture choices for production deployments, better handling of complex temporal patterns, and improved cold-start performance for new products or markets with limited historical data
TFT Interpretability - End-to-End Implementation
Description:
Implemented comprehensive interpretability suite for Temporal Fusion Transformer based on the original paper. Key features include: (1) Attention Weight Extraction & Visualization - Multiple plotting modes (mean, per-horizon, heatmaps) to understand which time steps the model focuses on. (2) Feature Importance via Variable Selection Networks (VSN) - Separate importance scores for static, past, and future covariates to identify which variables drive predictions. (3) Correlations & Time-Weighted Importances - Advanced metrics for debugging model behavior and building stakeholder trust. (4) Complete Documentation - Full official documentation with practical usage examples for real-world deployment.
Impact:
Transforms TFT from a 'black box' into an auditable, explainable model. Essential for production environments where stakeholders need to trust and understand AI decisions. Enables data scientists to debug underperforming models and communicate insights to non-technical teams
Research
PhD in Mathematics from École Normale Supérieure and published research
Uniform bounds for word length and group of bounded elements
PhD Thesis - École Normale Supérieure
My doctoral research was at the intersection of algebra, geometry and language theory, modeling mathematical structures as dynamic information systems.
The problem can be visualized from two complementary angles:
- The Language Lens (NLP): A group element is a "word" built from a given alphabet (generators). If we change the alphabet (a new "tokenization"), the word length generally changes drastically.
- The Network Lens (Graphs): The structure can be seen as a Cayley graph. Here, nodes are elements and edges are letters of the alphabet. Forming a word is like tracing a path from one node to another; the word length is the distance traveled.
The challenge: Usually, distance depends on the graph topology imposed by the chosen alphabet.
My contribution: I studied the existence of "intrinsic stability." I sought to identify the $G_{bound}$ subgroup, consisting of elements whose complexity (word or path length) remains uniformly bounded, regardless of how the graph is redrawn (change of alphabet).
Key results:
- I proved that this stability is impossible in negative curvature architectures (hyperbolic groups, where the subgroup is trivial).
- Conversely, in more regular structures (virtually abelian or nilpotent groups), I demonstrated that this subgroup precisely captures the finite information at the heart of the structure.
Transition to AI: This duality between structure (the graph) and information (the word) is at the heart of modern AI. Whether designing Graph Neural Network (GNN) architectures or optimizing NLP tokenization, I maintain this same mathematical approach: seeking robust invariants that survive data representation changes.
Publications
S-Transform with a Compact Support Kernel and Classification Models Based Power Quality Recognition
Journal of Electrical Engineering and Technology
February 2, 2022
In this paper, a novel method for power quality (PQ) events recognition is presented. Nine types of PQ events consisting of single and multi-stage disturbances are considered for study. For this task, features observed in the time frequency (t, f) plane have been used. Synthetic PQ events are generated using mathematical models. These signals are then projected in the time-frequency plane via the Stockwell Transform with a Compact Support Kernel (ST-CSK) providing the time-frequency resolution, energy concentration and robustness to noise. In this plane, PQ events are localized and characterized. The extracted features are then classified using several technics. The achieved results show that an overall accuracy of 100% has been obtained with Support Vector Machines and Random Forest classifiers even with signals embedded in high Additive White Gaussian Noise level (SNR=5dB). In the same conditions, XGboost classifier accurately detects 99.72% of PQ events.
Elements of uniformly bounded word-length in groups
Journal of the European Mathematical Society
May 17, 2019
This work introduces G_bound, the subgroup of elements with uniformly bounded word-length with respect to the change of generating set in finitely generated groups. Key results show that G_bound is a characteristic subgroup—finite for virtually abelian groups and trivial for non-elementary hyperbolic ones. Furthermore, for any finite group A, there exists an infinite group G such that G_bound is isomorphic to A. A variant dependent on the cardinality of generating sets is also analyzed.
Get in Touch
Let's discuss how I can help transform your AI initiatives into production-ready solutions
Availability
Currently available for freelance consulting projects. Typical response time: 24-48 hours.
Prefer email?
contact@yanisamirou.com