Nisrine BEL HADI
Machine Learning Engineer
4+ years exp
4+ projects
Python
SQL
Bash
Model Architecture & Training: 5/10
Research Depth & Innovation: 5/10
Data Pipeline & Feature Engineering: 4/10
Active 2 days ago
+33 (7) 58000327 Invite to interview
Message
Download CVCV
Overview
Technical skills
Timeline
Roles
Overview
An ML practitioner at a senior level with demonstrated ability to design and implement custom deep learning architectures for nontrivial problems. The strongest proven skill is custom neural-layer and architecture design, exemplified by the KANLinear class and the UKAN model in Unet_With_KANLayers.ipynb. There is limited evidence of production engineering practices such as testing, CI/CD, environment pinning and data/version management.
Technical skills
Python• Senior • 4y+ • 4 projects
SQL• Junior • 4y+ • 4 projects
Bash• Junior • 4y+ • 3 projects
Python
Requests
Flask• 4y+ • 4 projects
pySpark• 4y+ • 4 projects
Databases
Apache Kafka
Oracle• 4y+ • 1 project
PostgreSQL• 3y+ • 4 projects
MongoDB• 3y+ • 2 projects
ClickHouse• 2 projects
AI/ML
NumPy
Pandas
Scikit-learn
BERT
JAX
Flax
TF-Keras
Torchvision
Deep Learning
Jupyter Notebook
NLTK
Computer Vision
RAG
DevOps
AWS
CI/CD
Analytics
Matplotlib
Timeline
Université Paris XII - Paris-Est Créteil Val de Marne (Université Paris-Est)
Master's Degree •
Systèmes distribués et technologies de la data science (SDTS)
Ecole polytechnique d'Agadir
Other •
Engineering Degree – Data Science & Cloud Computing
Data & ML Engineer – Data Analytics (Internship, 3 months)
•
Junior
Ministry of Economy and Finance
•
Internship
Contributed to data analytics activities and machine learning support within a public sector environment. Assisted with data preparation and analysis workflows to support evaluation of results. Worked on delivery tasks over a short internship period focused on analytics outcomes.
Data Scientist – Deep Learning & Computer Vision (Internship)
•
Junior
3D Smart Factory
•
Internship
Developed and evaluated deep learning models for computer vision use cases. Worked with Kolmogorov–Arnold Networks (KAN) for model implementation and experimentation. Performed evaluation to assess model outputs for the task requirements.
Computer Vision
AI & Data Engineer – Machine Learning & RAG (Internship)
•
Junior
Teal Technology Services
•
Internship
Worked on machine learning tasks with a focus on retrieval-augmented generation (RAG) to improve performance. Contributed to evaluation and validation of model behavior. Applied engineering practices to integrate components into an AI workflow.
RAG
AI & Data Engineer – AI Systems & Machine Learning (Final-year Internship)
•
Junior
Opensee
•
Internship
Supported the development of AI and machine learning solutions from data preparation to model evaluation. Integrated multiple tools and AI capabilities to improve solution quality. Strengthened software engineering practices through deployment-oriented work on AI systems.
Senior AI/ML Engineer
Confidence: Medium ML Engineer
ML engineer at a senior-leaning level with a strength in prototyping custom neural architectures and algorithmic layers for applied ML problems. The strongest proven skill is implementing and integrating a novel B-spline-based KANLinear layer and KANLayer into a 3D segmentation model, as seen in the KANLinear and UKAN code. The public artifacts do not show production MLOps, formal experiment tracking, unit tests or deployment-ready serving code.
Model Architecture & Training
5/10
How well models are designed and trained
Custom model architecture and training code present, including a novel KANLinear implementation and end-to-end training loops in both PyTorch and TensorFlow.
Evidence
Data Pipeline & Feature Engineering
4/10
How data is prepared for models
Reasonable data pipelines and preprocessing for 3D medical volumes, tokenization and embedding steps for NLP, and augmentation scripts for class balancing.
Evidence
nisrine03/Kolmogorov-Arnold-Network/Unet_With_KANLayers.ipynb: functions preprocess and get_sub_volume and class BrainMRIDataset
nisrine03/Kolmogorov-Arnold-Network/IRM_Preprocesing_basic_model.ipynb: DataGenerator class and combine_masks_to_single_label
nisrine03/-Policy-Learning/Embedding_BERT.ipynb: tokenize_function and saving embeddings.npy
Experimentation & Evaluation
3/10
How results are measured and tested
Custom experiment loops, metrics and plotting are implemented but lack formal experiment tracking, reproducibility controls and systematic ablation.
Evidence
nisrine03/Kolmogorov-Arnold-Network/Unet_With_KANLayers.ipynb: training loop with DiceCoefficient metric and epoch logging
nisrine03/Kolmogorov-Arnold-Network/IRM_Preprocesing_basic_model.ipynb: model.fit usage with callbacks collecting losses and dice_coefficient
nisrine03/-Policy-Learning/RL_JAX.ipynb: RLAgent.train loop with printed stats and returns
MLOps & Deployment
2/10
How models are shipped to production
Minimal MLOps artifacts; models are saved to disk but there is no serving, versioning, monitoring or CI/CD evidence.
Computational Efficiency
3/10
How efficiently computing resources are used
Some performance-aware choices such as jitted JAX functions, DataLoader workers and device placement are present but there is no systematic profiling, quantization or GPU-memory optimization reporting.
Research Depth & Innovation
5/10
Depth of research and new ideas
Evidence of research depth via a custom B-spline-based linear layer, adaptive grid update and least-squares coefficient fitting, plus a custom RL environment and policy network.
Expertise
Computer Vision & Image Analysis• Senior
AI Agents & Agentic Workflows• Senior
Finance & FinTech AI• Middle
Technologies
Computer Vision
CI/CD
AWS
RAG
Torchvision
Flax
Recommendations
- Design and implement research-grade medical image segmentation prototypes that require custom layers and 3D architectures - extend and harden the KAN layer for experiments and ablations.
- Build RL-based feature selection and model-interpretability tools leveraging the JAX/Flax agent implementation and formalize evaluation pipelines for reproducible experiments.
- Develop NLP preprocessing and embedding pipelines into repeatable ETL modules and add experiment tracking (W&B or MLflow) and unit tests for reliability.
- Package trained models for inference with clear serialization, tests and a lightweight serving interface (TorchScript/ONNX/Triton or TF SavedModel) and add CI for reproducibility.
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Senior Data Scientist
Confidence: Medium ML Practitioner
An ML practitioner at a senior level with demonstrated ability to design and implement custom deep learning architectures for nontrivial problems. The strongest proven skill is custom neural-layer and architecture design, exemplified by the KANLinear class and the UKAN model in Unet_With_KANLayers.ipynb. There is limited evidence of production engineering practices such as testing, CI/CD, environment pinning and data/version management.
Statistical Rigor
3/10
Correct use of statistics
Some formal statistical checks and reporting are present (ADF test, classification reports), but there is limited uncertainty quantification, hypothesis testing discipline, multiple-comparison control, or deeper causal/robustness analysis.
Evidence
MASI_index/lstm.ipynb: adfuller stationarity test and ADF output
-Policy-Learning/tf_idf.ipynb: stratified train_test_split and classification_report
Data Wrangling & Cleaning
5/10
Preparing and cleaning data
Solid data ingestion and preprocessing pipelines for heterogeneous inputs are implemented - including MRI NIfTI loading, mask combination, sub-volume sampling and text cleaning/tokenization - though pipeline hardening and provenance/versioning are limited.
Evidence
Kolmogorov-Arnold-Network/Unet_With_KANLayers.ipynb: preprocess, get_sub_volume and BrainMRIDataset implementations
-Policy-Learning/Embedding_BERT.ipynb: clean_text and tokenize_function saving embeddings.npy
Exploratory Analysis & Visualization
4/10
Exploring and visualizing data
Notebooks include many visualizations and basic interpretation, but plots are used mostly to show results rather than to drive iterative hypothesis testing or to document rejected hypotheses.
Evidence
Kolmogorov-Arnold-Network/IRM_Preprocesing_basic_model.ipynb: show_slices and visualize_img utilities and loss/metric plots
MASI_index/lstm.ipynb: autocorrelation_plot and time series plots around stationarity checks
Predictive Modeling
5/10
Building models that predict
Demonstrates end-to-end model development across domains with bespoke architectures and training loops (3D U-Net variants, custom KAN layers, JAX RL agent, LSTM forecasting). Cross-validation, calibration and rigorous CV schemes are limited to basic train/test splits.
Evidence
Kolmogorov-Arnold-Network/Unet_With_KANLayers.ipynb: UKAN class, KANLinear implementation and training loop using DataLoader
-Policy-Learning/RL_JAX.ipynb: TextFeatureSelectionEnv, PolicyNetwork and RLAgent.train implementation
Business Insight & Impact
1/10
Turning analysis into business value
There is domain application context (medical segmentation and MASI forecasting) but little evidence of translating results into business metrics, cost-of-error analysis, or operational decision rules.
Evidence
MASI_index/README.md: describes MASI index forecasting objective
MASI_index/lstm.ipynb: LSTM forecasting pipeline for the MASI index
Reproducibility & Notebook Hygiene
3/10
Clean, repeatable analysis
Some reproducibility signals exist such as random_state usage, PRNG keys and model saving, but there is no environment pinning, CI, data/version management or systematic notebook-to-module refactor evident.
Evidence
Kolmogorov-Arnold-Network/Unet_With_KANLayers.ipynb: model.save('model_cnn.h5') and train_test_split(random_state=42)
-Policy-Learning/RL_JAX.ipynb: use of jax.random.PRNGKey for deterministic behavior
Industries
Health Care• Senior
Financial Services• Middle
Technologies
Deep Learning
Jupyter Notebook
Scikit-learn
JAX
Matplotlib
Pandas
NumPy
BERT
NLTK
TF-Keras
Recommendations
- Develop research or prototype systems that require custom neural layers and 3D medical image segmentation models.
- Implement proof-of-concept RL-based feature selection or token-pruning systems for NLP pipelines.
- Build time-series forecasting prototypes for financial signals where preprocessing and stationarity checks are needed.
- Contribute to model hardening and MLOps: add tests, reproducible environments, data/versioning and CI to productionize existing notebooks
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories:
Middle Backend Developer
Confidence: Medium Data Platform
Backend developer (Middle) specializing in real-time cryptocurrency streaming pipelines and analytics. The strongest proven skill is building streaming ETL and technical-indicator calculations using Spark Structured Streaming integrated with Kafka, as shown by spark_consumer.py and Kafka_producer.py. Production hardening, API contract design, auth/secret management and observability are not evidenced in the code.
API Design
2/10
How well APIs are designed
Minimal formal API design: the front end and services call a simple /stream endpoint without versioning, idempotency, or a consistent error contract.
Evidence
Dashboard.py: fetch_data requests.get('http://localhost:5000/stream')
flask_app.py: send_to_flask posts to FLASK_ENDPOINT ('http://127.0.0.1:5000/stream')
Data Layer & Database
3/10
Working with databases
Solid use of Spark Structured Streaming with an explicit schema and window functions to compute technical indicators, but no migration history, transaction boundaries, or hand-tuned queries/indexes are present.
Evidence
spark_consumer.py: defined StructType schema and casts for close/volume
spark_consumer.py: calculate_rsi / calculate_bollinger_bands / calculate_macd using Window functions
Scalability & Performance
2/10
Handling load and speed
Scalable components (Kafka, Spark) are used, but there is little production tuning or resilience: single partition defaults, no backoff/retry strategy, no connection/timeouts tuning or documented partitioning strategy.
Evidence
Kafka_producer.py: create_kafka_topic uses NewTopic(..., num_partitions=1, replication_factor=1)
spark_consumer.py: spark.readStream from kafka and df.writeStream.foreachBatch -> sends each partition synchronously to HTTP endpoint
System Architecture
3/10
Overall system structure
Reasonable service decomposition into producer, stream processor, model service and dashboard, showing an end-to-end pipeline design, but configuration, secret management and graceful degradation are not implemented.
Evidence
Kafka_producer.py, spark_consumer.py, flask_app.py and Dashboard.py present distinct producer, processor, prediction and UI components
Architecture pattern: Kafka -> Spark Structured Streaming -> HTTP ingestion -> Dash UI (multiple standalone scripts)
Security & Auth
1/10
Protecting data and access
Very limited security posture: hardcoded endpoints and brokers, no authentication/authorization, no input validation, and secrets/config are placed in code.
Evidence
Kafka_producer.py: KAFKA_BROKER and BASE_URL present as plain variables
flask_app.py and spark_consumer.py: FLASK_ENDPOINT hardcoded to 'http://127.0.0.1:5000/stream'
Reliability & Observability
1/10
Stability and monitoring
Minimal reliability and observability: print-based error handling, basic try/except, no structured logging, no metrics, no retries with backoff or timeouts, and no graceful shutdown handling.
Evidence
spark_consumer.py / flask_app.py: send_to_flask wraps requests.post in try/except and prints errors
Kafka_producer.py: delivery_report prints delivery status; no retry/backoff logic
Expertise
Messaging & Real-time• Middle
Python• Middle
Industries
Financial Services• Middle
Technologies
Python• Senior • 4y+ • 4 projects
Flask• 4y+ • 4 projects
Apache Kafka
pySpark• 4y+ • 4 projects
Requests
Recommendations
- Harden production readiness: add retries/backoff, request timeouts, structured logging, and metrics (Prometheus/OpenTelemetry) around HTTP and Kafka interactions.
- Implement a clear API contract and server routes with versioning and authentication for the /stream endpoint and centralize configuration and secrets (environment variables or External Secrets).
- Improve Spark/Kafka production settings: choose appropriate topic partitioning/replication, batch sizes, checkpointing, and test performance with realistic partitioning and state management.
- Add unit and integration tests plus CI to validate data flows and model inference behavior; add input validation for incoming messages and checks around model loading
Repositories
The developer's experience in this domain has been verified based on AI analysis of the following repositories: