Feb 2024 – Present · 2 yrs 3 mos

• Color and Font Variations System (Adobe Express Templates)
• Developed a generative system that creates visually aesthetic variations of Adobe Express templates by modifying color schemes and fonts across elements.
• Designed and implemented a complex end-to-end pipeline including preprocessing, post-processing, and scoring modules to ensure high-quality outputs.
• Fine-tuned an LLM-based model for style generation, with post-processing modules correcting errors and ensuring design consistency.
• Built a ranking/scoring module using LLMs to evaluate and prioritize variations based on aesthetic quality, enabling selection of the best designs.
• Query Understanding Model for Semantic Role Labeling
• Fine-tuned a 1B-parameter LLM to extract structured filters and values (e.g., created_by, modified_date, language, mime_type) from natural language queries in a document search system, enabling downstream semantic search.
• Led development of a synthetic dataset generation pipeline covering functional and non-functional variations, ensuring robust model training across diverse query patterns.
• Designed evaluation framework and golden dataset creation, addressing complex annotation challenges for large-scale query understanding.
• Collaborated extensively cross-team and cross-product, ensuring model integration, validation, and deployment within Adobe’s search ecosystem.
• Video Semantic Embedding Model (Adobe)
• Collaborated in a 3-person ML team to enhance Adobe’s video understanding system by fine-tuning state-of-the-art models (Gemma, LLaVA) on Adobe-acquired and generated video datasets.
• Built a robust evaluation pipeline using LLMs as judges to benchmark multiple approaches, achieving several percentage points improvement over prior systems.
• Focused on semantic video search across multiple dimensions including content, scene, style, camera, shot, action, emotion, and events.
• Contributed model fine-tuning, dataset preparation, and evaluation

Feb 2021 – Mar 2024 · 3 yrs 1 mo

• Video Understanding and Search (Adobe Express, Stock, and Firefly)
• Architect & Core ML Engineer
• Built a video understanding and retrieval pipeline powering semantic and hybrid video search across Adobe Express, Stock, and Firefly — supporting millions of user queries at production scale.
• Leveraged Adobe’s in-house CLIP-based multimodal model to jointly encode text and visual semantics, enabling contextually rich video discovery.
• Engineered a scalable video processing pipeline to extract and analyze key frames, followed by temporal self-attention–based aggregation to generate compact and discriminative video embeddings.
• Developed a comprehensive evaluation framework to benchmark video–query retrieval at scale — evaluating thousands of queries against curated datasets and producing standard metrics such as precision, recall, and F1-score.
• Optimized the system for high throughput, low-latency inference, and robust scaling across multimodal search workloads.
• Served as architect and core contributor, driving end-to-end design, modeling, and deployment of Adobe’s cross-product video understanding capabilities.
• Agentic Template Generation (Adobe Express)
• Architect & Core ML Engineer
• Contributed across ideation, architecture, and end-to-end execution of an AI-powered platform for automated promotional short video or still template generation using user products, images, and multimedia.
• Designed pipelines to analyze user intent and multimedia content via captioning and embedding models, providing structured input for downstream generation.
• Leveraged LLMs to generate video scripts, which were segmented into scenes with optimized layouts retrieved from a layout store.
• Built modules to fit text, images, and video assets into scenes automatically, producing polished, end-to-end video outputs.
• Served as architect and core contributor, driving pipeline design, LLM integration, layout optimization, and scalable production deployment.

Feb 2018 – Feb 2021 · 3 yrs

• ML Training Platform (Adobe Sensei)
• Architect & Lead Engineer
• Architected and developed a scalable ML training platform for Adobe Sensei, enabling data scientists to launch GPU/CPU-based training and data processing jobs seamlessly via Jupyter notebooks or CLI.
• Integrated elastic file systems and dynamic resource allocation for optimized utilization of compute and storage.
• Designed APIs to export standard training metrics to platforms such as TensorBoard and Amazon S3, supporting unified experiment tracking.
• Built and integrated on-demand Spark clusters for large-scale data preparation, fully accessible through the platform’s notebook and CLI interfaces.
• Drove platform adoption across ML teams by combining usability, scalability, and cost efficiency within a unified training infrastructure.
• Content Processing Framework (Adobe Sensei)
• Founding Team Member · Architect & Core Contributor
• One of the founding members of a 10-person team that built Adobe’s Content Processing and Inference Platform, powering large-scale AI/ML and content services across Adobe products.
• Architected and implemented a highly scalable, multi-mode inference framework supporting sync, async, batch, streaming, and long-running jobs for workloads like video and content processing.
• Designed a JSON-defined DAG–based orchestrator to chain AI/ML services, content processors, and I/O modules with seamless data and control flow across distributed systems.
• Developed critical data connectors integrating with Adobe’s internal storage ecosystem to ensure reliable, high-throughput content movement.
• Engineered key features for reliability, scalability, and performance, conducting deep dives into Redis and HBase to optimize fault tolerance, caching, and horizontal scaling for high-load environments.
• Served as both architect and hands-on engineer, driving system design, performance optimization, and platform adoption across Adobe’s AI infrastructure.

Dec 2015 – Feb 2018 · 2 yrs 2 mos

• Marketing Campaign Platform (Adobe Internal Project)
• Primary Architect & Core Engineer
• Led architecture and end-to-end development of a large-scale marketing automation platform enabling Adobe marketers to define user profiles, rule-based segmentation, and multi-channel campaigns.
• Designed and implemented a highly scalable rule engine using Drools, powering real-time decisioning on explicit user activities (clicks, visits) and implicit behavioral events (cross-product usage within time windows).
• Built core infrastructure leveraging AWS SQS, SNS, and DynamoDB, ensuring high throughput, fault tolerance, and horizontal scalability.
• Collaborated with an 8-member cross-functional team to deliver the product from concept to production, driving personalization and engagement automation at scale.
• Help Content Recommendation (Adobe – Photoshop)
• Machine Learning Engineer / Applied Scientist
• Developed an intelligent help content recommender system for Photoshop to surface contextually relevant tutorials and videos based on in-product feature usage.
• Designed a hybrid recommendation pipeline combining mixture model–based user clustering with stochastic collaborative filtering, improving relevance and personalization for diverse user segments.
• Built a click- and view-driven ranker to optimize tutorial ordering for engagement while maintaining content diversity.
• Implemented an anomaly detection module to identify feature usage sequences correlated with user drop-off, enabling proactive retention insights.

Dec 2014 – Dec 2015 · 1 yr

• Rule based Personalization, Recommender Systems

Jul 2014 – Dec 2014 · 5 mos

• Adobe Core Technologies.