AI / ML / DL Engineer · Data Scientist Building production-grade ML systems — evidence-grounded LLMs, calibrated quantile models, hybrid decision engines — with the MLOps surface (observability, supply-chain, drift) that keeps them honest.

📫 marwabensalem30@gmail.com · 🌐 linkedin.com/in/marwa-ben-salem-573465b6

Evidence-first LLM pipeline with hallucination guards, seniority-calibrated outputs, prompt versioning, drift detection, and a $/resume cost meter.

Stack: FastAPI · Streamlit · Pydantic v2 · Redis · arq · Mongo · FAISS / Qdrant · OpenAI · OTel · Prometheus · Helm · Docker Highlights: LLMProtocol + VectorStoreProtocol abstractions · 1,326 tests · 7-job CI (Trivy · SBOM · perf regression gate) · retrieval eval with recall@k / MRR / nDCG + bootstrap CIs · multi-pod-safe JWT revocation · Redis-backed daily cost circuit-breaker Live demo: resumeforge-bg29.onrender.com

Hybrid 5-signal weighted scorer + bounded LLM reasoning layer. Locked weights, append-only audit log, intentionally-inert evaluation pending real outcomes. Determinism formally verified to 1e-9 precision across local, CI, and production.

Stack: Pydantic v2 · sentence-transformers · OpenAI gpt-4o · MongoDB Atlas · Streamlit · Docker · GitHub Actions · Hugging Face Spaces Highlights: Architecture-as-Contract (immutable design doc + execution rules) · 261 hermetic invariant tests in 1.10s · 3-job CI (privacy audit · tests · auto-deploy to HF) · decision_trace with counterfactual sensitivity replays + dominant-signal coding · Protocol-based Mongo / LLM / embedding abstractions · frozen 3-environment determinism proof artifact Live demo: huggingface.co/spaces/MarwaBS/job-decision-engine

4-model comparison (XGBoost · LightGBM · Random Forest · multi-task PyTorch TabNet) on 4,504 cleaned NYC listings. Honest R² = 0.815 — documented the removal of the PRICE_PER_SQFT leakage that inflated v1 to R² = 0.997.

Stack: XGBoost · PyTorch · scikit-learn Pipelines · Optuna · SMOTE-ENN · SHAP · FastAPI · Streamlit · DVC · MLflow Highlights: Macro F1 = 0.724 with per-class threshold tuning · SHAP explainability · fairness-by-borough analysis (diagnostic) · assert_no_leakage() enforced in CI · 64 tests / 88% coverage · Dockerfile multi-stage + Trivy + SBOM + Dependabot Live demo: docker compose up --build (or try the Streamlit app locally)

End-to-end data science pipeline on BLS OEWS + Census microdata. Multi-quantile XGBoost (P10/P50/P90) returns a calibrated uncertainty range — point-estimate R² is the wrong frame on a truncated cohort.

Stack: XGBoost quantile · FastAPI · Streamlit · Prometheus · Redis · Hugging Face Spaces · GHCR Highlights: 149 tests · 3-job CI (matrix 3.11+3.12) with Trivy scan gating GHCR push · APIKeyHeader auth + env-driven CORS + proxy-aware slowapi rate limiting · cluster-wide drift monitor · scheduled retraining workflow Live demo: huggingface.co/spaces/MarwaBS/high-pay-salary-predictor

Languages: Python 3.12 (primary) · SQL · Bash ML: scikit-learn · XGBoost · LightGBM · PyTorch · Optuna · SHAP LLM / AI: OpenAI · LangChain · FAISS · Qdrant · Sentence-Transformers · custom LLMProtocol abstraction Serving: FastAPI · Pydantic v2 · Streamlit · Uvicorn · arq · Redis · Mongo Infra: Docker (multi-stage) · Kubernetes / Helm · Hugging Face Spaces · Render · GHCR · docker-compose Observability: OpenTelemetry · Prometheus · structured JSON logs · locust perf regression CI / DevEx: GitHub Actions · ruff · mypy (strict) · bandit · pytest · Trivy · CycloneDX SBOM · Dependabot

• Sharpening interview prep across the 4 target lanes.
• Open to Senior AI Engineer · ML / DL Engineer · ML Systems Engineer · Data Scientist roles.
• Based remote/open to relocation.

_{Last updated: 2026-04-19 · Repos green CI-verified on HEAD · MIT-licensed}