Constraint
Noisy retrieval
Search across large codebases was difficult to trust without a repeatable evaluation harness.
ML systems engineer · Paris
I build ML systems
for messy data, tight constraints,
and real users.
I’m Yassine Bentayfor. I work across retrieval, optimization, and data platforms, building the evaluation and infrastructure needed to take models beyond a notebook.
Available Sept. 2026 · CDI / permanent
Dual-degree engineering student at ENSIIE × EMI, based in Évry-Courcouronnes.
Working languages: Arabic, French, and English.
Working methodConstraint → decision → result
How I approach a system
Start with the
failure mode.
I identify what can fail, choose the smallest defensible intervention, and define how I’ll know whether it worked. The tools come after that.
Decision
Separate recall from precision
Use a bi-encoder for retrieval, ColBERT for reranking, and benchmark each stage rather than treating RAG as one black box.
Evidence
+47% nDCG
Retrieval quality improved while test coverage reached 86% and p95 runtime fell by 41%.
Oracle Labs · professional/internal work · summarized to respect confidentiality
Selected workFour systems / four constraints
Selected projects
Four systems chosen for different problems: relevance, safety, numerical accuracy, and distributed orchestration. Each one includes the decision behind the result.
ProfessionalSearch
Building a code-search stack I could measure end to end.
A production search stack combining ingestion, embeddings, FAISS retrieval, ColBERT reranking, LLM summaries, and a parametrized evaluation suite.
+47%retrieval nDCG
86%test coverage
−41%p95 runtime
Implementation notes
Constraint
Relevance gains were hard to attribute to the embedding model, retriever, reranker, or summary layer.
Decision
Benchmark E5-Large-v2, mxbai, and ColBERT separately, then validate the complete path with parametrized Pytest coverage.
Scope
Professional/internal work. Architecture and metrics are intentionally summarized; no proprietary corpus or implementation is exposed.
Applied MLClinical
Keeping hard safety constraints out of the language model.
A portion-level MILP meal-planning engine coupled to guardrailed recipe generation, deterministic validators, and auditable JSON artifacts.
99.2%feasibility
<1 ssolve time
120stress tests
Implementation notes
Constraint
Weekly plans had to satisfy hard nutrition and allergy constraints while remaining varied, deterministic, and auditable.
Decision
Keep constraint satisfaction inside MILP and restrict the language model to presentation, behind allergy guards and regression checks.
Scope
Feasibility is an engineering result across 120 stress tests. It is not a claim about clinical outcomes.
ReproductionResearch
Reproducing the paper before optimizing the implementation.
Neural-network implied-volatility modelling under Black–Scholes and Heston, reproducing Buehler et al. (2019) with a Dockerized TensorFlow-GPU workflow.
1.55e−8BS MSE
1.14e−6Heston MSE
1.1 Mtraining samples
Implementation notes
Constraint
Heston implied-volatility labels are expensive: pricing and root-finding must be accurate before the network can learn anything useful.
Decision
Generate Black–Scholes samples with Latin hypercube sampling and Heston samples with COS pricing plus Brent inversion, then reproduce the paper’s figures.
Scope
This is a reproducibility and engineering project, not a claim of a novel pricing method.
Personal buildPipeline
Turning a million accident records into a reproducible pipeline.
Accident data moves from Kafka and MinIO through Spark and Trino to OpenSearch, with Airflow coordinating the complete workflow.
1,065,052rows processed
6service stages
1orchestrated DAG
Implementation notes
Constraint
Raw accident files were too large and inconsistent for a one-off notebook or a manually repeated analytics workflow.
Decision
Separate ingestion, object storage, transformation, query, indexing, and orchestration so every stage can be rerun and inspected independently.
Scope
The row count establishes scale; the project’s real subject is service boundaries and reproducibility.
Also built
Time series / Paris
Smart City Traffic Forecasting
ARIMA vs Prophet vs LSTM · Airflow · PostgreSQL · 17% RMSE improvementRetrieval / French + English
Bilingual RAG Evaluation
E5-Large-v2 · semantic vs sliding-window chunking · FastAPIExperienceFive roles / 2023—2026
Where I’ve worked
My roles have covered model research, cloud infrastructure, search, clinical systems, and governed data. In each one, I worked on making experimental systems dependable.
-
Feb — Aug 2026Procter & Gamble
Data Engineering Intern
3 silent drifts caught before production impactPySpark and Delta Lake pipelines on Azure Databricks, with schema, completeness, freshness and drift gates plus Unity Catalog governance.
-
Jun — Sep 2025Integrative Phenomics
Data Science / Applied ML Intern
22% less clinical iteration timeRe-platformed a fragile R/Shiny prototype into a Python–React system with biomarker analytics, LLM summaries, and MILP nutrition optimization.
-
Jul — Sep 2024Oracle Labs
Machine Learning Engineer Intern
+47% retrieval quality; 86% test coverageBuilt and evaluated the semantic-search path from embedding benchmarks through FAISS retrieval and ColBERT reranking.
-
Jan — Jul 2024Thales Group
Cloud Engineering Apprentice — MLOps
Manual handoff moved to zero-touch deploymentEvent-driven AWS architecture with Lambda, EventBridge, SNS, least-privilege IAM, containers, CI/CD, reusable IaC and structured logging.
-
Jun — Sep 2023Oracle R&D
R&D Intern — AI Code Assistance
+17% factual alignmentStarCoderPlus 15B with LoRA, custom RAG, FAISS, AlignScore/BLEU evaluation, and syntax-validity checks for SuiteScript generation.
Technical focusLinked to project evidence
Technical strengths
Grouped by the problems I used them to solve. Each section links back to the relevant work.
Find and rank the right context
Embeddings, FAISS, ColBERT, RAG, semantic chunking, nDCG, precision, golden sets and ablations.
Used in semantic searchTurn requirements into systems
MILP, deterministic validators, CPLEX/CBC, FastAPI, PostgreSQL, guardrails and auditable artifacts.
Used in the clinical optimizerBuild the data underneath
PySpark, Delta Lake, Kafka, Airflow, Databricks, Trino, MinIO, quality gates and governed history.
Used in the distributed pipelineMake the path repeatable
AWS, Azure, Docker, Terraform, CI/CD, least-privilege IAM, observability and reproducible evaluation.
Used across professional rolesCompetitions and certifications
Other results
Kaggle House Prices
of 5,000+ entries · 2025E-Toufoula Hackathon
7 engineering-school teams · 2024Think AI Morocco
400+ applicants · 2024ODC CP Hackathon
Morocco / Africa · 161 teamsFour more distinctions
- Oracle OCI Generative AI Professional · 2024
- Oracle OCI Foundations Associate · 2024
- IBM DevOps Professional · 2024
- NASA Space Camp Commander’s Cup · 2019
Available September 2026
I’m looking for my next
ML engineering role.
I’m interested in ML engineering, data, and applied-science roles where evaluation and production quality actually matter.
bentayfor.yassine@gmail.com