Nauman – LLM, Python, AI agent development

The AI Productivity Platform for In-House Legal.

• Built AI Revise V3, achieving an approximate 60% win rate across most evaluation datasets.
• Designed and implemented the end-to-end AI Revise evaluation system, from metric computation through an LLM-based judge.
• Contributed to Project Re: Invent, rebuilding the company's core product around an agentic AI foundation.
• Shipped the Workflow Engine for Re: Invent, enabling users to author complex workflows in simple, natural language.

Test automation for mobile and web applications.

• Led the ML team by setting technical direction and driving delivery across multiple machine learning projects.
• Deployed an optimized Visual Regression API pipeline serving 14M+ requests at approximately $1,200/month using Knative on GKE.
• Developed and trained an in-house transformer-based OCR model from scratch, reaching >95% character recognition across 30K+ Japanese and English characters.
• Built LLM-powered tools for test creation and test generation, producing tests, Gherkin, and Playwright code from feature PRDs.
• Implemented a distributed reinforcement learning algorithm for automated web navigation.
• Trained a semantic segmentation engine using U²Net for estimating visual changes in web screenshots across multiple versions of the same web page.
• Optimized the parameters of a NodeJs based application using genetic algorithms.
• Built an OpenCV-based visual difference engine for web screenshots.

Machine learning and computer-vision services.

• Implemented lyrics extraction with EAST and Tesseract for a music-sheet parsing program.
• Improved a proprietary information-retrieval algorithm for parcel images captured by cameras.
• Bootstrapped infrastructure for a SaaS product that extracted information from receipts.
• Trained BERT-based classifiers on datasets of 100K+ examples and used GPT-2 and T5 for unstructured-text parsing.
• Deployed large language models in production with Docker, Cloud Run, and Kubernetes.
• Performed analysis on a regex/rule-based system.
• Compared performance and accuracy of a rule-based system with a deep learning NLP and vision system.
• Upgraded rule-based components to machine-learning-based data-driven components.
• Designed new machine-learning-based components to satisfy business needs.
• Deployed models cost effectively on modern cloud platforms including AWS Lambda.
• Built an internal data analytics platform to identify patterns to help grow business.
• Extracted and organized unstructured OCR data from shopping receipts.
• Trained deep learning models using BERT and Spacy to extract useful information from receipts data.
• Deployed scalable deep learning models on GCP using Cloud Run and Cloud Functions.

A research Project for Fish Biodiversity Protection

• Implemented a forward-backward motion tracker using OpenCV for fish tracking in camera images.
• Developed a UI for debugging and annotating the fish-tracking algorithm and its ground truth.