https://suryal.dev/ Long-form engineering and architecture notes on enterprise AI, MCP and agentic operations, SRE/platform engineering, and cloud modernization. en-us suryal.k90@gmail.com (Suryaprakash Lakshmanan) Sun, 20 Jul 2026 00:00:00 +0000 https://suryal.dev/articles/inside-the-forward-pass.html https://suryal.dev/articles/inside-the-forward-pass.html Sun, 20 Jul 2026 00:00:00 +0000 Internals #1: the LLM forward pass taken apart stage by stage on a real quantized model with MLX on Apple Silicon — tokenization, embeddings, attention, logits, and the generation loop, every claim verified bit-for-bit against the library's own kernels. Five animated stages, runnable on any Apple Silicon Mac. https://suryal.dev/articles/mcp-summit-mumbai-2026.html https://suryal.dev/articles/mcp-summit-mumbai-2026.html Tue, 17 Jun 2026 00:00:00 +0000 Two days, 16 sessions, one through-line: multi-agent AI doesn't fail inside the protocol — it fails at the seams between agents, tools, and protocols, and every failure is a missing contract. A practitioner's field notes from MCP Dev Summit Mumbai 2026. https://suryal.dev/articles/principal-propagation-mcp.html https://suryal.dev/articles/principal-propagation-mcp.html Sat, 14 Jun 2026 00:00:00 +0000 Giving an AI agent MCP access to a regulated PostgreSQL database — carrying the end user's verified identity to the data source, with authentication, authorization, and audit enforced there. Built with MCP, OIDC/Keycloak, and Postgres 18 SASL OAUTHBEARER. https://suryal.dev/articles/mlx-breaking-point.html https://suryal.dev/articles/mlx-breaking-point.html Fri, 12 Jun 2026 00:00:00 +0000 A reproducible benchmark pushing local LLM inference from 3B to 123B parameters on an Apple M5 Max with 128GB unified memory — decode, TTFT, the real memory ceiling, and the 14-inch thermal tax, measured with MLX. https://suryal.dev/articles/sap-northstar-ai-architecture.html https://suryal.dev/articles/sap-northstar-ai-architecture.html Sat, 06 Jun 2026 00:00:00 +0000 A practitioner's breakdown of SAP's AI-Native North Star Architecture, covering context engineering, HANA's evolution, Knowledge Graph, MCP/A2A, data lake concerns, ECC/Z-code, and AI governance.