Souhila M.

I'm an AI Research Engineer with 7+ years of experience building fast, trustworthy ML replacements for expensive computations, neural surrogates for physics solvers, reliability-gated prediction systems, and the evaluation infrastructure that validates them. My core differentiator: I don't just build fast models, I build models that know when they're wrong. Every surrogate includes a confidence gate that flags uncertain predictions and falls back to the original solver, so you never get a silent wrong answer. What I Build: Simulation Acceleration : I replace your slowest computations (FEM, CFD, combustion, optimization solvers) with neural surrogates that deliver predictions in milliseconds instead of minutes. Reliability-Gated AI Systems : Every surrogate comes with a confidence layer: softmax entropy monitoring, bootstrap ensemble disagreement, MC-dropout uncertainty quantification. High confidence uses the surrogate; low confidence triggers solver fallback including safety-gated continuous learning that correctly rejected a catastrophic retrain before deployment. Evaluation & Research Engineering :I design evaluation harnesses that separate real improvements from noise: multi-seed validation (810-run Monte Carlo with statistical power analysis), root-cause failure classification, and honest trade-off reporting. Optimization & Decision Systems :MILP formulation from scratch (YALMIP, intlinprog), genetic algorithms, PSO, hybrid metaheuristics. I've distilled complex optimization solvers into lightweight neural policies with three-tier reliability cascades for real-time deployment, including edge hardware (ONNX export). Technical Stack: - Python: NumPy, SciPy, PyTorch, JAX,for scientific computing - MATLAB & Simulink: dynamic system modeling, control design, solver pipelines, Excel VBA integration - Simulation & CAE: ANSYS (3D FEM structural analysis), OpenDSS (power systems), Simulink real-time - Industrial: PLC programming, SCADA/HMI (TIA Portal), predictive maintenance systems - Numerical Methods: FEM, ODE/PDE integration, spectral analysis, convergence analysis - ML for Engineering: Bayesian-regularized ANNs, delta-learning multi-fidelity surrogates, imitation learning (DAgger), reinforcement learning (DQL) - Evaluation: Latin Hypercube DOE, Monte Carlo validation, Cohen's d power analysis, confidence calibration Selected Results: - 440x speedup: MILP solver distilled into a neural policy for multi-UAV task allocation, 44% fewer collisions than the teacher solver (p<0.001, 810 Monte Carlo runs) - 92.27% accuracy: combustion physics surrogate from 484-configuration Latin Hypercube campaign, with per-output confidence gates - 50x power reduction: RIS-aided MIMO beamforming optimizer replaced by dual-role ANN - Safety-gated learning: continuous learning loop that correctly rejected a bad retrain before deployment Credentials: - IEEE published: ISPA 2022 (DOI: 10.1109/ISPA54004.2022.9786315) - Validated by Postwork AI (Silicon Valley): designed a novel DQL prototype, 44+ hours of recorded expert reasoning used as AI training data - Cross-domain delivery: combustion physics, structural mechanics (ANSYS FEM), telecommunications, energy systems, autonomous control - Upwork Top Rated Plus, 100% Job Success Score Remote, async-first, well-documented delivery. Available for project-based and retained engagements.