Chandan

Machine Learning Engineer with 1.5+ years building production Document AI, Computer Vision, and Generative AI systems; developed LLM-powered document pipelines processing 50K+ documents/month and deployed FastAPI ML services on Azure. Delivered a 40% inference memory reduction for Mistral-7B, engineered OCR-free table extraction with 93% layout accuracy, and implemented HITL workfows, AI guardrails, and ClearML-backed experiment traceability to ensure reliable production operations.

Software Trainee Engineer - Visionet Systems Pvt. Ltd. - India

(2023-02 - 2023-07)

• Evaluated and benchmarked multiple transformer architectures for unstructured document processing, comparing accuracy, inference latency, and deployment trade-offs to inform production model selection.
• Optimized enterprise ML service debugging workfows, reducing troubleshooting time by 8% through data preparation, pipeline validation, and production MLOps practices.

AI/ML Engineer - DocVu.AI - Visionet Systems Pvt Ltd - India

(2023-08 - 2024-09)

• Engineered end-to-end Document AI pipelines for enterprise mortgage processing, streamlining document classification, information extraction, indexing, and document-level Q&A for production workfow
• 40% inference memory reduction on Mistral-7B. Fine-tuned for domain-specific Document VQA; applied grouped-query attention (GQA) and sliding-window attention (SWA) to make 100+page document inference tractable without scaling GPU spend.
• 93% table extraction accuracy, OCR-free. Engineered a transformer pipeline (TATR / PubTables-1M, DONUT, DESSERT) that replaced a brittle OCR chain failing on complex multi-page mortgage documents. Benchmarked on internal test sets, deployed to production, monitored for layout drift.
• Zero-downtime async pipeline. Designed FastAPI microservices with thread-based parallel execution orchestrating extraction, indexing and Q&A, non blocking, thread-safe, production-hardened.
• 30% reduction in deployment downtime. Versioned model artefacts on Azure Blob Storage with rapid rollback eliminated the "redeploy and hope" cycle, enabled safe A/B model transitions in production.
• Full experiment traceability across 3+ parallel development tracks. Managed all runs, checkpoints and model versions in ClearML any result reproducible or rollable-back on demand.

Master of Science (MS) - Artificial Intelligence & Machine Learning - University of Birmingham (2025-12)