Vishal M

Transformed enterprise AI infrastructure by consolidating 15+ ML microservices into a unified Kubernetes-based LLM inference platform, reducing infrastructure costs by 60% and accelerating release cycles from days to hours. Designed an LLM output-to-OCR coordinate grounding framework with multi-line highlighting to improve document field localization accuracy from 30% to 85% for production document intelligence. Built AI platform observability using Prometheus/Grafana and performance validation with k6, and optimized inference with structured extraction, confidence scoring, SLA validation, and partial document processing. Engineered an automated LLM evaluation & benchmarking framework using Python and Elasticsearch/ELK, and delivered a production-grade Prompt Management Platform with SDK, REST APIs, RBAC-enabled UI, versioning, environment management, tagging, and atomic rollback. Led backend architecture and enterprise search/deduplication pipelines processing 100K+ documents daily wi

Designed a scalable distributed ML inference engine using Python, Celery, Redis, Kafka, Kubernetes, and AWS ECS, serving 20+ models with 99.9% availability and improved classification performance. Led migration of inference workloads from AWS ECS to Kubernetes, improving scalability, resilience, and resource utilization while reducing infrastructure costs by 35%. Architected an event-driven inference platform with Kafka/Celery using a time-windowed consume-and-republish approach to eliminate transaction loss and enable seamless horizontal scaling. Built a large-scale knowledge ingestion pipeline with AWS Lambda, Dagster, Elasticsearch, FastAPI, and S3 to process 50TB+ data and enable 10x faster search and retrieval. Developed checksum-based deduplication pipelines to remove ~40% redundant content and generate structured datasets for LLM training, owned Elasticsearch search/autocomplete services with caching and weighted relevance, and served as SME for Kibana/ELK monitoring for proacti

Developed and deployed KYC document classification models for Aadhaar, PAN, Passport, and other identity documents using Google Vertex AI to automate document categorization in enterprise verification workflows. Improved inference performance through model quantization and batch processing, achieving nearly 3x faster prediction latency while maintaining production-grade accuracy.