At TenX, I developed predictive models for credit-card upsell and cross-sell, using customer, transactional, and product features. I applied SHAP analysis and tree-based feature importance to improve explainability and increase the initial hit rate from approximately 50% to over 90%.
I also addressed data drift in newer mammography data for a breast cancer detection model, improving recall from 62% to 85% during evaluation at a leading cancer hospital in Pakistan. I optimized the Python backend and made annotations clearer for clinical interpretation.
At TenX, I built an end-to-end Text-to-SQL pipeline that ranked client Vertica tables by query relevance and supplied selective context to an LLM. I set up local GPU inference using vLLM and quantized Gemma and Qwen models, with validation guardrails and safe-failure mechanisms.
Previously, at AIO, I led a team project that automated sentiment analysis of more than 10,000 restaurant menu-item feedback entries, reducing manual analysis time by 80%. My projects also include a diffusion-model tool for refining character sketches and a Flask/FastAPI application for interactive image masking.
