Selected Portfolio

Case Studies in Computational Inference & Neural Architectures.

A rigorous examination of applied machine learning, focusing on scalability, interpretability, and the intersection of optimization and deep learning.

CLINICAL NLP EVALUATION

PARHAF CliBench

A benchmark for measuring how 7B-9B language models handle structured information extraction on real French clinical notes.

Why it matters

Hospitals need reliable extraction for de-identification, infection tracking, and structured reporting. This benchmark tests whether compact open models are actually ready for those workflows.

Technical Challenge

Clinical IE requires more than medical knowledge: models must emit valid structured outputs, preserve exact spans, and stay robust across heterogeneous tasks and note styles.

Methodology / Evaluation

Evaluated 6 instruction-tuned LLMs and GLiNER2 on 4 PARHAF tasks, covering 5,005 documents and 65,065 predictions, with micro-F1, schema conformity, empty-output rates, latency, and bootstrap confidence intervals.

terminal Repository description Technical Report play_circle Demo

REALTIME AUDIO TRANSCRIPTION

Asifonix

A low-latency ASR system thought for multiuser real-time applications in controlled environments.

Why it matters

Performant ASR tools exist, but sensitive environments like healthcare and other regulated industries, cannot rely on third-party APIs. Deploying reliable real-time transcription on-premise often means building from scratch.

Technical Challenge

Delivering stable, incremental text across concurrent sessions without the latency or privacy tradeoffs of cloud-hosted alternatives.

Methodology / Evaluation

Three-layer on-premise system: web client, FastAPI gateway with admission control, and a Faster-Whisper gRPC backend. Validated under load from 5 to 500 concurrent users.

LICENSE PLATE RECOGNITION

GetThePlate

End-to-end automatic license plate recognition system based on deep learning.

Why it matters

Few reliable systems handle the specificities of Algerian license plates. The contribution is both the Deep Learning models and a complete software suite (rest API and mobile app), ready for field deployment.

Technical Challenge

No public dataset existed for Algerian plates. Building a reliable end-to-end pipeline meant going from data collection to model fine-tuning and serving.

Methodology / Evaluation

Custom dataset of 2,500 augmented images gathering, labelisation, quality and representativity evaluation. Computer Vision Model trainig and evaluation, MLOps and model serving.

terminal Repository