Salman

Python Engineer

Binary • Temps plein

Dec 2025 - Present • 5 mos

• Spearheaded the transition from Flask to FastAPI, introducing async capabilities, Pydantic models, JWT-based auth, and modular dependency injection—boosting API response time and maintainability for AI/ML services. • Built a workflow orchestration engine using Celery and RabbitMQ, powering user-defined pipelines (via drag-and-drop UI) with state tracking and dynamic execution logic for data and ML flows. • Integrated and productionized MLflow for full model lifecycle management, enabling reproducible training, tracking, and model serving for AutoML, GPR, XGBoost, and Random Forest models. • Fine-tuned LJSpeech-based speech-to-text and text-to-speech pipelines using Hugging Face’s facebook/mms-tts-eng and PyTorch; handled audio preprocessing, tokenization, and training loop optimization. • Deployed LLM inference pipelines (Hugging Face Transformers) for text generation, summarization, and intent classification—optimized for both real-time API and batch processing, reducing latency by 30%. • Implemented end-to-end conversational AI modules, including semantic intent recognition and summarization models, with RabbitMQ-Celery integration for autonomous handling of document and chat streams. • Evaluated and integrated foundation models (e.g., GPT-like) for multi-turn conversation understanding, contributing to internal agentic workflows and context-aware response generation. • Fine-tuned a pretrained speech-to-text model (facebook/mms-tts-eng) using LJSpeech dataset for custom voice synthesis tasks; implemented data cleaning, tokenization, and training loop using PyTorch and Hugging Face’s transformers and datasets libraries. • Integrated fine-tuned models into workflow engines with RabbitMQ and Celery, enabling autonomous processing of document/text streams in production.