1. Shrew Built an AI-powered low-code browser automation platform using Playwright with visual workflows, session recording, and AI powered self-healing execution to handle UI changes, broken selectors, and runtime failures. 2. ShrewVoice Built a voice-driven automation system integrating ElevenLabs, MCP, and Playwright to execute Shrew-generated workflows via natural speech with deterministic browser execution. 3. RCM Automation Built a HIPAA-compliant end-to-end RCM automation pipeline using UiPath to identify follow-up claims, automate payer portal navigation, retrieve claim status, generate LLM summaries, and update systems, processing 2,500+ claims and reducing a one-month workflow to two days. 4. Voice Utilities + MCP Integration Integrated ElevenLabs conversational AI with an MCP server for natural voice interactions in water utility services, enabling consumption tracking, usage insights, and outage reporting. 5. Nexus - Account Overlap Platform (POC) Developed a full-stack AI platform to detect and manage account overlaps across company and partners, enabling scoring, escalation, and resolution of high-value sales and partnership opportunities. 6. CEQA Automation Platform (POC) Built a FastAPI-based AI system using OCR, NLP, and Gemini to classify CEQA documents (Exempt, ND, MND, EIR), generate justification reports, flag environmental risks, and provide an interactive UI with status tracking, maps, caching, and chatbot assistance. 7. Clinical Trial Graph AI (POC) Engineered a Neo4j + FastAPI + Gemini platform for ADC clinical trials enabling natural language-to-Cypher queries, visualization of PK/safety metrics (AUC, Cmax, AE grades), and AI-driven insights via AWS-hosted dashboards.

Worked with Users Team in ZohoDesk | Developed a Retrieval-Augmented Generation (RAG) system using FAISS for document retrieval | Occasionally suggested new ideas to the Zia Team | Built production-grade JavaScript and ReactJS features

Worked with the Speech Profiling team | Learned about audio feature extraction | Conducted audio classification using CNNs, achieving peak accuracy | Addressed challenges with varying sampling rates | Gained valuable insights into audio engineering and feature extraction

Developed a Fewshot Doc2Vec Model | Began research with ResNet-50 and transitioned to DiT, a vision transformer pretrained on a document dataset | Extracted document embeddings and performed few-shot training using Triplet Loss | Achieved 100% similarity search accuracy with well-labeled data | Discovered insights that classification models sometimes perform like few-shot models and vice versa | Conducted multimodal training by combining visual features from ResNet and textual features from BERT, resulting in strong generalization | Mentored by Fasil Saidalvi