experience

ZS Associates · Toronto, Canada

• Designed and built LangGraph agents that author CMC (Chemistry, Manufacturing & Controls) regulatory documents for a global pharmaceutical client — improving output correctness while cutting authoring time by 50%.
• Built RAG systems and agent tools that ground generation in source material, so agents produce accurate, context-appropriate content.
• Architected scalable AWS ECS services to isolate compute-heavy authoring and document-mapping workloads, keeping them independent from the core application.
• Designing an agentic HAQ (Health Authority Questions) feature — LangGraph agents that draft responses to regulator questions raised after CMC submissions.

VaultAI · Vancouver, Canada

• Architected an AI-powered client intake portal for private credit firms, enabling natural-language analytics on financial data and automated extraction from unstructured PDFs.
• Engineered an autonomous Due-Diligence Agent that orchestrates multi-source web research to synthesize credit risk and legal compliance into production-ready PDF reports.
• Designed a Multi-Company Financial Comparison Agent to benchmark fiscal performance across entities, reducing manual comparison effort by 60%.
• Built secure RAG pipelines for 100+ confidential documents, ensuring enterprise-grade privacy and reliable grounding for complex inquiries.
• Developed automated auditing workflows (Financial Health Checks, Bluesheet generation) using Supabase and Edge Functions for real-time processing.
• Fine-tuned Llama models with QLoRA for domain terminology — a 15% accuracy improvement with reduced inference latency.

Gilbert Research Center · Coimbatore, India

• Applied Pandas, NumPy, and OpenCV to run ETL on 10K+ medical X-ray images, with data-quality checks and normalization pipelines for deep learning.
• Trained CNN models in PyTorch on medical imaging datasets, achieving 87% classification accuracy for early-stage diagnostics.
• Implemented data augmentation and cross-validation, improving robustness and reducing overfitting by 12% on test data.