Senior ML/NLP Engineer for On-Device Financial Notification Parsing (Edge AI / TFLite)

We are building a next-generation personal finance tracker for the global market. The core feature relies on parsing banking SMS and push notifications to build an automated dashboard of bank accounts, credit cards, and recent transactions. Because global bank formats change constantly and vary by language, traditional Regex is not scalable. We need a highly accurate (90-95%+) Machine Learning model to handle this dynamically. The Objective Train a custom Named Entity Recognition (NER) model or fine-tune a Small Language Model (SLM) to extract specific financial data from unstructured notification text. To comply with strict Play Store privacy policies, this model must run 100% offline on the user's Android device. Required Extracted Entities: Bank / Credit Card Name Account Number (Last 4 digits) Currency & Amount Transaction Type (Debit / Credit) Merchant / Vendor Name Date and Time Scope of Work & Deliverables - Synthetic Data Generation: Due to privacy laws, you will not be provided with real user SMS data. You must script and generate a highly diverse dataset of 50,000 to 100,000 synthetic banking notifications covering worldwide formats using LLMs (OpenAI/Claude). - Model Training: Train a robust NLP model (e.g., MobileBERT, TinyBERT, or a similar architecture) on this BIO-tagged dataset. - Optimization & Conversion: Apply Knowledge Distillation and INT8 Post-Training Quantization to convert the trained PyTorch/TensorFlow model into a .tflite (LiteRT) format. - Strict Size Constraint: The final .tflite file MUST be under 50 MB so it can be packaged directly into our Android assets folder without heavy cloud downloads. - Output: The model must accept a raw string (the notification text) and return a clean, structured JSON object. Required Skills - Deep expertise in Python, NLP, Named Entity Recognition (NER), and Hugging Face. - Proven experience converting large models to TensorFlow Lite for mobile deployment. - Strong understanding of model quantization (INT8) and edge computing.