Replaced Whisper API-based intent inference with a fine-tuned, quantized SLM deployed on-device, cutting latency by ~3x and pushing intent classification accuracy beyond 98% while eliminating external API costs

Designed end-to-end system architecture for the SLM pipeline on AWS EC2 including data ingestion, fine-tuning, quantization, serving, and MLflow-based experiment tracking with shadow deployments

Led AI R&D for the voice intelligence initiative by defining technical roadmap, setting latency SLAs, designing for horizontal scaling, and mentoring engineers across model development and evaluation workstreams

Fine-tuned Stable Diffusion with LoRA adapters for few-shot visual concept learning, reducing required training samples from thousands to under 20 while preserving generation quality

Designed a Pipecat agentic pipeline to serve custom TTS models via API with real-time audio streaming, turn-taking logic, and fallback routing under latency constraints

Built an agentic customer support system with dual-corpus RAG using ChromaDB, achieving a RAGAS score of 0.85 and ~40% ticket deflection rate orchestrated via LangGraph

Enhanced ML data pipelines for fraud-login detection using XGBoost, improving model precision and recall on production traffic

Integrated monitoring KPIs into Grafana dashboards for real-time visibility into model performance and data drift

Managed ML pipeline optimization and performance monitoring for production systems

Optimised 5+ ETL workflows using AWS Glue and PySpark, producing analytics-ready datasets at scale for downstream ML model training

Automated CI/CD pipelines with GitHub Actions and Jenkins, achieving 80% build/test coverage and boosting development efficiency by 25%

Built an object detection system with YOLOv8 achieving 92% accuracy on a custom automotive dataset and developed interactive Tableau dashboards integrated with AWS Data Catalog

Researched and implemented CNN and Transformer architectures for out-of-distribution (OOD) detection in autonomous driving datasets using PyTorch and co-authored a peer-reviewed publication

Built a custom NER pipeline using BERT and RoBERTa, improving entity extraction accuracy by 28%

Executed 2D sensor fusion via Python-based Kalman filter and implemented autonomous path planning with CARLA and ROS

Accelerated pose tracking by 40% via GPU-optimised optical flow techniques

Automated training and evaluation pipelines, reducing iteration cycles by 35%

Built stereo and IP camera calibration setups with OpenCV