Thirteen years across cloud, data, and enterprise systems — now focused on agentic enablement at production grade. I design the governed, auditable paths that let teams wire LLMs and agents into backend systems: MCP server frameworks, tool-use SDKs, and the middleware that absorbs identity, data governance, and guardrails.
Defined an MCP → federated-identity → enterprise-systems topology, now the org-standard pattern for agentic workloads. First proven across SAP HANA, OData, and SaaS landscapes — cutting new integration build time from weeks to days.
AI EnablementShipped an MCP framework with shared SDK libraries that absorb federated auth, query governance, and security guardrails — so consuming teams inherit production-grade safety by default instead of rebuilding it per system. Pioneered the FastMCP runtime in production.
AI EnablementArchitecting AIOps across three agentic domains — Monitoring & Alerting (triage, Dataproc/BigQuery monitoring, vulnerability tracking), Operations (change coordination, patch-day automation, system-health analysis), and Knowledge & Query (L1 knowledge and troubleshooting agents).
AIOpsProductionized an agent giving SRE and platform teams natural-language access to system health, sessions, and workload analytics on the HANA platform — now the reference implementation other agents are built from.
AIOpsDrove FinOps governance delivering verified, recurring cost savings — BigQuery slot reservations, GCP committed-use discounts, rightsizing, and quota management — with no performance degradation.
Cloud & FinOpsBuilt Terraform provisioning frameworks that eliminated ~80% of manual ClickOps and self-service platforms that offloaded ~70% of L1/L2 tasks — plus CLI-generated, production-ready MCP servers and an interactive docs site that turns one-off builds into a repeatable authoring path.
PlatformArchitected Apache Spark / Dataproc platforms processing 200M+ events per day — PySpark pipelines over Pub/Sub streaming and batch ingestion at multi-terabyte scale, with GKE-based orchestration and capacity monitoring.
Cloud & DataSide-of-desk research, run in public — benchmark harnesses, live dashboards, and evals. Code on GitHub, every number reproducible; the write-ups live in the logs.
A benchmark harness for local LLMs on Apple Silicon (MLX): a model ladder from 3B to 123B on a 128GB M5 Max — decode, TTFT, the real memory ceiling (spoiler: not where Apple says), and the 14-inch thermal tax. Live dashboard, every number reproducible from the repo.
Next in the series: can a local model pick the right tool for an operations alert? The number that actually matters for autonomous ops — eval design in progress, same build-in-public rules.
The little bot in the corner of this page. Scripted telemetry by default — but click load real brain and it boots a ~0.6B local model on your own GPU via WebGPU. No server, no API key, nothing leaves your machine.
This site's thesis as a runnable lab: an AI agent reaches a Postgres banking database carrying the end user's verified identity — not a shared service account — with authentication, authorization, and audit enforced at the data source. Role-based PII masking, per-tool RBAC, and a Postgres-18 path where the database itself denies unauthorized columns.
The platform layer for agentic workloads — governed, auditable, and reusable across teams.
Multi-cloud platforms with a deep Google Cloud foundation — provisioned, shipped, and operated as code.
Big-data platforms supporting streaming and batch ingestion at multi-terabyte scale.
Observability, chaos engineering, and InfoSec collaboration for audit-ready systems.
Leading AI & agentic platform engineering (MCP) and AIOps — plus large-scale GCP migrations, Spark/Dataproc data platforms, FinOps governance, and SRE.
Ansible-based automation for patch management; SAP/Linux hardening for audit mandates; HA/DR with Pacemaker and SAP HANA clusters for Fortune 500 clients.
System automation and DR setups for global SAP clients; contributed to the HANA Center of Excellence on performance tuning and migration assessments.
A working library of engineering and architecture writing — enterprise AI, MCP & agentic operations, platform engineering, and cloud modernization. The long-form lives here; the shorter takes go out on LinkedIn.
The LLM forward pass taken apart and verified bit-for-bit on a real model — five animated stages
Field notes from MCP Dev Summit Mumbai 2026 — every failure is a missing contract
Why brittle RPA is giving way to agentic AI — agents that read documents by meaning, call ERP and procurement tools through controlled interfaces, and escalate only true exceptions. Scripts for stable tasks, agents for variable workflows, humans for judgment boundaries.
Read on LinkedIn ↗ inLinkedIn · PostA defense of the Model Context Protocol: standardized auth, auditing hooks, and guardrails for tool access where no clean API exists. Context bloat isn't MCP's fault — it's architecture. Atomic, focused tools stay surgical.
Read on LinkedIn ↗If you're working on agentic platforms, MCP, or AIOps — or just want to compare notes — say hello. Email is the surest way to reach me.
suryal.k90@gmail.com