AI Engineer
Infergience Quotient Pvt. Ltd.
As an AI/ML Engineer at Inferigence Quotient, I owned end-to-end development and deployment of production ML systems across computer vision, audio, and NLP. Multimodal Emotion Analysis Pipeline: Designed and deployed a real-time system fusing visual, audio, and text signals via weighted late fusion, served through a production REST API. Fine-tuned HSEmotion for facial emotion recognition, Wav2Vec2 for audio, and RoBERTa for text classification. Built a streaming ingestion pipeline handling 5–10s audio/video chunks with off-the-shelf VAD and noise suppression, integrated over WebRTC. Deployed on AWS using Docker and Kubernetes, with custom Prometheus metrics for production monitoring. Object Detection & Few-Shot Learning: Trained and fine-tuned YOLO variants and Faster RCNN for surveillance under low-data, real-world conditions. Led R&D on a few-shot object detection methods with custom loss functions, achieving a +17.23 AP point improvement over baseline. Separately implemented Multi-Similarity and Coherence loss functions on Faster RCNN. Automated Drone Landing Pipeline: Built a multi-stage vision pipeline - runway detection, pose/parameter regression (reproducing a research paper with no reference implementation), EKF, and PID control - integrated with PX4 autopilot and validated in X-Plane 11 simulation. MLOps Tooling: Extended MLflow (internal tool "iremat") by adding Datasets and Overview tabs to improve experiment tracking and model comparison workflows for the team. Stack: PyTorch, AWS (EKS), Docker, Kubernetes, WebRTC, MLflow, Prometheus.
