Harsh Sanas

About

AI/LLM engineer focused on shipping production-grade agents that automate support, sales intake, and ERP-connected workflows. Stands out for combining strong orchestration and guardrails with measurable business outcomes, including 45% faster support handling, ~$1.2M annual savings, 18% higher customer satisfaction, and 99.5%+ reliability in production.

Key Strengths

• Shipped production LLM agents with 99.5%+ uptime and sub-2-second latency
• Delivered measurable business impact: 45% reduction in support handling time
• Generated approximately $1.2M annual support cost savings
• Improved customer satisfaction by 18%
• Strong reliability engineering for LLM systems under API rate limits and latency spikes
• Designed robust fallback and confidence-based escalation to human agents
• Architected auditable multi-step agent workflows with structured outputs and state management
• Built schema-validated, guardrailed integrations that prevented malformed data from breaking downstream systems
• Data-driven evaluation and monitoring across offline tests and live production metrics
• Improved live reliability by reducing fallback rates 30% through prompt and classifier iteration
• Handled messy ERP integrations with normalization, data quality controls, and idempotent writes
• End-to-end ownership of Python backend services for real-time telemetry systems
• Designing robust state-reconciliation and idempotency to handle out-of-order/duplicate OCPP events
• Building and operating Kubernetes deployments with Helm, probes, autoscaling, and zero-downtime rollouts
• Implementing GitOps CI/CD with GitHub Actions + ArgoCD and fast rollback via Git revert
• Planning and executing cloud-to-on-prem migrations using blue-green cutover and CDC with zero downtime
• Designing Kafka streaming pipelines for bursty high-frequency data with ordering and deduplication controls
• Improving ML forecasting reliability via better preprocessing and minimum-history thresholds

Experience

Education

Skills

Languages