About Me
Experience
Core Skills
Technical Solutions
Industry Domains
Personal Work/Writings
MLUI Mobile: Autify OCR vs. Google OCR
Comprehensive performance comparison of Autify's in-house OCR system against Google Cloud OCR and EasyOCR, with detailed methodology, results analysis, and sample evaluations. Autify OCR achieved 91% accuracy on mobile screenshot text recognition.
Token Compression: Reducing Attention Waste?
Explored using LLMs to compress multiple tokens into single tokens for more efficient transformers. Demonstrated that 2048 hidden dimensions can compress ~8 tokens losslessly using a two-stage encode-decode architecture with LoRA fine-tuning.
Long Pythia
Explored token length extension when it was common to have 2048 or 4096 context length.
Machine Learning Features in Autify for Mobile
Comprehensive overview of AI-powered mobile testing features including Visual Regression Testing (VRT), Visual Self-Healing algorithms, and the upcoming Visual App Explorer (VAX) for autonomous app navigation and bug discovery.
Solving Automated App Navigation: A Use-case
Detailed exploration of behavior cloning techniques for automated mobile app navigation, comparing regression vs heatmap approaches, and demonstrating how U²Net successfully models uncertainty in tap location prediction.
Applying Modern Deep Learning in Autify
Comprehensive overview of deep learning applications in software testing, including visual regression detection, genetic algorithm optimization, graph neural networks for HTML analysis, and reinforcement learning for intelligent test discovery.
Getting the Most Out of Pre-trained Models
Deep dive into pre-trained NLP models like GPT-2 and T5, exploring their capabilities for text generation, question answering, summarization, and transfer learning applications. Originally published on Toptal.
Recent Advancements in AI
Comprehensive overview of AI breakthroughs in 2019, covering text generation, image synthesis, audio creation, and video/animation technologies that were transforming industries.
This Icon Does Not Exist — GAN for Icon Generation
An application of Generative Adversarial Networks to icon generation, exploring mode collapse and training challenges with a custom dataset.
Deploy ML on Cloud Run
Complete tutorial on deploying machine learning models to Google Cloud Run using Docker and GPT-2 as an example.
Cloud Run — Future Tech
Explored Google's revolutionary serverless container platform and its comparison with traditional serverless and container technologies.
Deep Learning in Cloud
Explored the cloud computing options for deep learning.
Professional Experience
Autify
Bootstrapped project Lexa
- Generate PRD from Source Code
- Vibe-coded entire desktop app using cursor in Swift UI
- Vibe-migrated the desktop app to Electron for multiple platforms
Bootstrapped project Genesis
- Designed Prompts for the following:
- Generate Test Cases from PRD
- Generate Test Scenarios from Test Cases
- Generate Playwright code from Test Scenarios using Agentic AI
- Build Backend/frontend using LLMs
Developed Step Suggestion Chrome Extension
- Wrote frontend code in PReact
- Wrote Backend in Python/FastAPI
- Developed the algorithm to slim down the HTML
- Designed the prompt to provide next step suggestions based on current page
Core Deep Learning Work
- Developed MLUI, an ensemble of ML models for parsing screenshot of mobile apps
- A custom OCR model based on encoder-decoder transformer architecture trained on synthetic data
- Retina net fine-tuned on rico + internal dataset for detecting UI elements
- A post processing model to merge the output of the above two models
- Deployment of MLUI at scale on GKE using cost-effective GPU & queue optimizations
- Explored Reinforcement Learning for automated testing
Built the baseline ML Infrastructure at Autify
- Set up Google Cloud GKE for ML Deployment
- Set up pipeline for VRT (Visual Regression Testing) algorithm deployment at scale on GKE
- Fine-tuned Adobe Semantic Web Segmentation model
- Explored Visual Self-Healing using Template Matching
Early Exploration of Application of ML on Web Apps
- Implemented OpenCV-based web screenshot comparison
- Fine-tuned BERT on HTML data
- Optimized the weights of an internal feature extraction algorithm using evolutionary algorithms