tanweerulhaque

1) Config-driven ML Workflow Orchestrator : - Crafted a configurable centralised MLOps library to automate end-to-end ML workflows on GCP using Airflow. - Orchestrated GCP VM provisioning, git repository builds, Docker executions, hyperparameter tuning, real-time Slack monitoring, GCS artifacts versioning, logging and cleanup across workflows in parallel. Automated the full CI / CD pipeline. - Reduced model iterations from 2 weeks to 1 day. 100+ production-ready models within 2 months by multiple teams. - Scaled framework to support all 7 internal DS pipelines, successfully deploying risk models with 60–80% precision. 2) Real-time Affluence Estimation Engine : - Engineered a real-time service with tree-based models to estimate user affluence in pre-onboarding using sparse multi-source data (profile, bureau, apps, AA transactions, etc) to reduce low-margin acquisitions, leading to org-wide adoption. - Accelerated model iterations by 1–2 weeks by implementing an extendable gRPC data fetching library from backend DBs. - Scaled service to 8M+ users (9 req/s, 5s average latency). Enabled Redis caching to keep batch and real-time scores in sync. - Impacted 70% acquisitions, boosted financial runway by 1 year and improved KPI metrics (RGU, Avg funds added, etc). - Leveraged PySpark batch pipelines to automate model scoring and deliver actionable insights. Achieved a 12x surge in salary users (2K/month). Unlocked high ROI targeting for Credit cards, Savings account campaigns and Personal loan affinity users. 3) DS On-Call MLOps Monitoring : - Owned 24/7 primary on-call for 12 critical DS services and pipelines for 60% of a 4-year tenure, maintaining 99% uptime. - Drove cross-functional system stability by productionising 35+ PRs, resulting in 3x cost reduction in GCP, AWS and K8s.

1) Anti Money Laundering (AML) Modelling in Batch and Real-time : - Devised a foundational algorithm to detect ”Fast-in-Fast-out” mule activity, powering 25 heuristics and 3 models. - Productionized a Transformer model with 40% precision. Increased DS recall by 15% to 40. Runs on 400M transactions daily. - Integrated 3 real-time tree models on pre-onboarding and other stages with 60% precision and 400 ms average latency. - Built a self-learning case prioritisation service to auto-block 500+ daily fraudsters, cutting ops workload by 4x. - Maintained a 20K+ feature pipeline and drove brainstorming sessions to contribute new categories based on domain feedback. 2) In-House Liveness and KYC Video Fraud Detection : - Architected a multi-modal ensemble Deep Learning system for video KYC liveness and fraud detection. - Developed speech, frame-quality, lip-sync, spoof, and OTP signal to compute a unified risk score using a curated dataset. - Replaced a major external vendor, saving 120K dollars in vendor API costs and Reserve Bank of India penalties. - Deployed to evaluate 20M+ users over three years with zero errors and downtime. Achieved an overall system precision of 75% from 30%, along with standalone spoof detection growing from 40% to 95%.

Translated FDD and TUT requirements into 10 robust test scripts and unit tests. Tracked defects via Jira and transitioned the manual testing workflows into streamlined pipelines resulting in 20% reduction in QA effort.

Instrumented ETL pipelines for 17 US FHIR resources and analysed 2.5M nested JSON records. Mastered PostgreSQL CRUD operations (10x faster) with asyncpg and multiprocessing, achieving 3x faster real-time data processing.