Work Experience
Senior AI Engineer – Justdial
Jan 2024 – Present · Bengaluru, India (On-site)
-
Optimized inference for large language and vision models like Llama3.1 8B, Llava1.5v 7B, and Llama3.2-vision 11B using Nvidia's Triton Inference Server. Integrated TensorRT backend and KV cache to reduce latency by 74.12% (from 17s to 4.4s). Currently implementing Flash Attention and refining autoscaling with Ray Serve.
Tech Stack: TensorRT-LLM, Nvidia Triton Inference Server, PyTorch, FastAPI, Docker, HuggingFace Transformers
-
Designed and deployed a system to dynamically display profile images based on user-searched categories. Used fine-tuned SBERT, MLflow for tracking and deployment, and Milvus for free-text vector search. Improved CTR by 16.37% within a month.
Tech Stack: SBERT, Milvus, Llama3.2-vision 11B, OpenCV, Guardrails
-
Developed an on-premise video transcoding pipeline for HTTP Live Streaming (HLS), reducing costs by avoiding cloud services.
Tech Stack: RabbitMQ, Docker, FFmpeg, Nvidia NVENC unit
-
Led the design and deployment of a centralized MLflow system to support reproducibility, collaboration, and model lifecycle management across teams.
Tech Stack: MLflow, AWS S3, MySQL, Nginx-Proxy server
-
Built an image enhancement pipeline using GAN models (RealESRGAN_x4plus, RealESRNET_x4plus, GFPGAN), integrated with AWS S3. Enhanced over 1M images, improving overall user experience.
Tech Stack: GANs, AWS S3
-
Designed and deployed image and video metadata extraction system, processed over 100M images and 1M+ videos across major Indian cities.
Tech Stack: Docker, Flask, Gunicorn, MongoDB, MySQL, Apache Kafka, Redis, Go, Python, MediaPipe, ImageMagick
Data Engineer Intern – CropIn
Nov 2023 – Jan 2024 · Bengaluru, India (On-site)
-
Led deep learning efforts for plant disease prediction using ResNet9-based models across multiple agricultural datasets.
-
PlantVillage: Achieved 99.40% validation accuracy on a clean dataset; test accuracy dropped to 5% on real-world images due to limited background variation.
-
PlantDoc: Managed class imbalance in a 2223-image dataset; reached 52.34% validation accuracy but only 27.50% test accuracy on real images.
-
FieldPlant: Achieved 73.50% validation accuracy from 5348 imbalanced images; test accuracy was 17.25% on real-world images.
-
Blended Dataset: Created blended images using Kaggle and random backgrounds, achieving 99.96% validation accuracy and 70.80% test accuracy.
-
Worked on improving generalization of models for real-world deployment; targeted test accuracy uplift from 70.80% to 85%.
-
Used advanced model tuning techniques like dropout, gradient clipping, L2 regularization, data augmentation, normalization, and weighted loss.
-
Gained hands-on experience with deep learning, data imbalance handling, and deployment challenges in agricultural AI.
