SATHISH JAYARAMAN

Designed and built production-grade platforms entirely through AI-assisted development — zero hand-written code. Architected event-driven systems with Apache Airflow DAGs, Redpanda (Kafka) streaming, and multi-agent LLM orchestration using LangChain with RAG-augmented reasoning via ChromaDB. Built ML pipelines with XGBoost/LightGBM for predictive analytics and autonomous resolution through specialist AI agents. Developed SaaS platforms using Next.js, TypeScript, Supabase (Auth + Postgres + RLS), Daytona SDK for sandboxed execution, and OpenRouter for LLM routing — featuring WebSocket RPC hot-config, multi-channel integrations, and encrypted credential vaults. All development powered by Claude Code, Cursor, TaskMaster AI, and Ralph Loop.

Architected hybrid Edge-Fog-Cloud platform processing real-time OT data from manufacturing facilities worldwide using Kubernetes, Kafka, TimescaleDB, and Spark. Designed distributed system achieving low-latency edge analytics with cloud-based historical analysis. Implemented fault-tolerant pipelines with cross-region replication ensuring zero data loss during regional outages. Platform maintained high availability while handling substantial YoY data growth, enabling customers to reduce manufacturing downtime and improve product yield through predictive analytics.

Resolved critical performance bottleneck by migrating from Celery to Apache Airflow with Kubernetes Executor and Spark integration. Legacy system had limited concurrency for model training with lengthy deployment cycles. Implemented dynamic Kubernetes pod scheduling with Spark-based distributed training, achieving dramatic improvements in training throughput, substantially reduced model deployment time, and optimized compute costs through spot instance usage and right-sized resource allocation. Enabled real-time adaptation of predictive models to changing manufacturing conditions.

Resolved severe production crisis where time-series query latencies degraded dramatically, making manufacturing dashboards unusable. Identified that Elasticsearch's inverted index and JVM garbage collection were unsuitable for high-frequency numeric time-series workload. Executed zero-downtime migration to TimescaleDB using phased approach with dual-write strategy, custom Spark-based historical data backfill, and gradual canary deployment. Achieved substantial query performance improvements, increased write throughput, significantly reduced storage footprint, and eliminated database-induced outages entirely.

Designed dual-schema database architecture resolving conflict between high write throughput from thousands of edge connectors and low read latency for real-time dashboards. Single schema couldn't efficiently serve both workloads. Solution separated data into write-optimized tables (minimal indexing, columnar storage, partitioned by edge connector) and read-optimized tables (aggressive indexing, continuous aggregates, partitioned by sensor), connected via Spark streaming ETL with exactly-once semantics. Achieved substantial write throughput improvements, dramatically reduced query latency, enabled scaling to significantly more concurrent users, and reduced anomaly detection time critical for preventing costly pharmaceutical batch losses.