Ai Developer

Built an end-to-end LLM pipeline for K-12 curriculum and assessment generation using agentic workflows (multi-step, prompt-engineered generation with validation loops) and post-trained custom models; generated content averaged 99%+ on an independent multi-dimensional quality evaluation spanning factual accuracy, curriculum alignment, difficulty calibration, and distractor quality. Post-trained LLMs via Direct Preference Optimization (DPO) with QLoRA (parameter-efficient fine-tuning) for constrained educational content generation — fill-in-the-blank, MCQ, and MSQ formats across 3 difficulty tiers — improving constraint adherence by 4.5% over the SFT baseline. Deployed and operated production ML services on AWS: model serving on SageMaker, containerized microservices on ECS/ECR, event-driven pipelines with Lambda + SQS, API Gateway with Elastic Load Balancing, S3 artifact storage, and CloudWatch monitoring. Fine-tuned open-weight Arabic LLMs using multiple SFT strategies, outperforming top Open Arabic LLM Leaderboard (OALL) models on multiple benchmarks in internal evaluations; co-authored a research paper on the methodology.

Architected and deployed a production-ready Vision-Language Model(VLM) pipeline featuring custom in-house text encoder, x8 quantized, and llama.cpp-powered inference engine (LLaMA C++), achieving real-time performance on Samsung Galaxy S25 devices. Utilized llama.cpp’s quantization and runtime optimization for MLLM deployment, engineering grammar-based shared objects that enable efficient GPU execution, achieving up to 3× throughput speedup. Built and optimized an on-screen AI model for a multi-class classification task, achieving 97.1% accuracy using Custom LightWeight classifiers based on convolution(CNN) and Transformer. Developed a custom lightweight CLIP model with 40% model size reduction via knowledge distillation and pruning, maintaining 90% of original accuracy. Applied it for a high-quality Sticker Search system. Designed and implemented a dataset preparation pipeline using web scraping and synthetic data generation, improving dataset size by 200%+. Incorporated advanced algorithms like LoRA on Stable Diffusion. Proposed and commercialized a novel approach for Barcode and QR code detection in mobile screenshots using synthetic data generation and text localization, achieving 92.7% accuracy on real-world datasets. Created an AI-powered WatchCrop solution that generates optimal wallpaper crops for samsung Smartwatches, achieving 92% user satisfaction through a custom Saliency Model.

Contributed to developing a Phonetic Search Feature to retrieve similar-sounding Indian names from user inputs, supporting 7+ languages. Spearheaded the creation of a diverse phonetic-focused dataset of Indian names, enhancing pronunciation-based search functionality by 80% and improving overall user experience. Gained working knowledge of Neural Phonetics models, including seq2seq architectures with LSTM and Time Distributed layers, to better align dataset design with model requirements. Successfully deployed the model on-device for real-world commercial applications(Phonetic Search Feature).