LA Tech Week 2026 — Google for Developers Codelabs & AI Workshops

Thu, 15 Oct 2026 14:00:00 +0000

Workshop Overview

At LA Tech Week 2026, Google for Startups and Google for Developers hosted hands-on codelab tracks for enterprise founders, engineers, and technical leaders. Alongside fellow Google Developer Expert Kiruthika Subramani, I co-instructed two interactive technical workshops detailing how to build production-grade agentic architectures with Google Cloud and the Model Context Protocol (MCP).

Workshop 1: Location Intelligence ADK Agent with MCP Servers

Interactive Presentation & Teleprompter:
Official Google Codelab:
Core Technologies: Google Agent Development Kit (ADK), BigQuery Geospatial Data, Google Maps Grounding Lite MCP, Gemini 2.5 Flash

Architectural Focus

Agent Development Kit (ADK) Core: Structuring autonomous agents with schema-enforced tool interfaces, state persistence, and error-recovery loops.
BigQuery MCP Integration: Bridging spatial data warehouses with natural language queries—enabling dynamic SQL generation, bounding box filtering, and spatial joins over enterprise logistics tables without manual query plumbing.
Google Maps Grounding Lite MCP: Validating place details, computing multi-modal travel times, and grounding model answers against live physical coordinates.
Production Guardrails: Rate-limiting spatial queries, handling disconnected MCP sessions, and sanitizing agent reasoning steps.

Workshop 2: Command & Control: Gemini, MCP & Agentic CLI

Interactive Presentation & Teleprompter:
Official Google Codelab:
Core Technologies: Antigravity CLI (agy), Developer Knowledge MCP, Agent Skills, Nano Banana 2 Image Gen, Google Cloud Run

Architectural Focus

Headless Agent Orchestration: Driving terminal-first agentic developer loops with Antigravity CLI (agy), managing multi-step project scaffolding and debugging sessions.
Developer Knowledge MCP: Exposing local codebase indexing, symbol trees, and architecture specs to agent runtimes without manual context dumping.
Multimodal Generation & Deployment: Chaining multimodal synthesis via Nano Banana 2 for visual assets directly into automated container packaging and deployment to Google Cloud Run.
Composable Agent Skills: Packaging reusable capability modules across teams to enforce testing, security, and deployment standards across agent prompts.

Reception & Industry Feedback

“Google for Startups and Google for Developers take over #LATECHWEEK… Thanks to my GDEs Rabimba Karanjai and Kiruthika Subramani for providing such wonderful and thought driven codelabs.” — Marty Fortier, Google Developer Ecosystem / Startup Lead

“LA TECH WEEK by a16z was hands-on. We built small tools around real questions from our day-to-day, like which truck a load needs and how to read proof-of-delivery paperwork faster, and got a map agent running in the cloud in a guided codelab… Thank you to Rabimba Karanjai and Kiruthika Subramani, our instructors in the Google for Developers codelabs.” — Elizabeth Moscoso, Founder & Logistics Operator

OpenXLA Summer DevLab 2026 — Benchmarking vLLM on XLA Backends

Thu, 25 Jun 2026 18:00:00 +0000

Talk Overview

Invited by Google’s ML and TPU teams to present at OpenXLA Summer DevLab 2026 hosted at Google Corporate HQ in Mountain View, CA, alongside speakers from Google, NVIDIA, and the Meta PyTorch team.

The talk details our empirical evaluation of the matured vLLM + OpenXLA stack, contrasting it with the native vLLM + CUDA serving path under identical enterprise workloads.

Key Technical Insights

1. The Runtime Collision: Dynamic CUDA vs. Static OpenXLA

GPU / CUDA: Inherently dynamic, memory allocation and kernel launches scale on the fly to arbitrary prompt lengths with near-instantaneous container startup.
TPU / OpenXLA: Built around high-density Systolic Arrays streaming deterministic matrix multiplications. XLA compiles rigid machine code ahead of time, requiring incoming variable-length prompts to be mapped into static padding buckets (128, 256, 512, 1024 tokens), introducing compute padding overhead.
The Compilation Wall: Uncompiled sequence configurations or cold starts trigger ahead-of-time compilation passes that can pause server execution for 20–30 minutes, complicating elastic autoscaling.

2. The Throughput Trap vs. Usable Goodput

Standard evaluations optimize for Peak Throughput (tokens/sec), masking degraded latency.
LLM inference splits into:
- Prefill phase: Compute-bound; dictates Time to First Token (TTFT).
- Decode phase: Memory-bandwidth-bound; dictates Time Per Output Token (TPOT).
We evaluate capacity using Usable Goodput — requests per second successfully meeting both TTFT ≤ 1000 ms and TPOT ≤ 50 ms simultaneously.

3. Interconnect Limits & Tail Latency Collapse

On single-chip configurations (e.g., Llama-3.1 8B), the system scales predictably, keeping P99 response loops under 38 ms.
For partitioned 70B models requiring inter-chip all-reduce tensor collectives, operating over baseboard PCIe Gen5 without NVLink bridges quickly saturates the bus. P99 decode tail latency spikes to ~466 ms/token, causing usable goodput to collapse from 9.3 down to 2.4 requests/second under peak concurrency.

4. The 2026 Unified vLLM-XLA Stack Architecture

Consolidated Control Layer: Google’s internal JetStream stack was officially archived in early 2026 and folded into the JAX-native tpu-inference engine, with vLLM providing the OpenAI-compatible frontend orchestrator.
Torchax Lowering: Standard PyTorch graphs are caught on-the-fly and lowered into StableHLO structures without eager Python interpreter bottlenecks.
Ragged Paged Attention v3 (RPA v3): Implemented in Google’s low-level Pallas and Mosaic kernel layout engines, RPA v3 enables dynamic ragged tiling inside HBM, fusing KV-cache scatter steps into the attention block and delivering up to 86% Model Bandwidth Utilization (MBU) during generation.

Resources & Artifacts

Presentation Slides:
Benchmark Harness & Test Suite:
Deep-Dive Engineering Post:
Recorded Session Stream:

Invited Talk | Rabimba Karanjai

LA Tech Week 2026 — Google for Developers Codelabs & AI Workshops

Workshop Overview

Workshop 1: Location Intelligence ADK Agent with MCP Servers

Architectural Focus

Workshop 2: Command & Control: Gemini, MCP & Agentic CLI

Architectural Focus

Reception & Industry Feedback

OpenXLA Summer DevLab 2026 — Benchmarking vLLM on XLA Backends

Talk Overview

Key Technical Insights

1. The Runtime Collision: Dynamic CUDA vs. Static OpenXLA

2. The Throughput Trap vs. Usable Goodput

3. Interconnect Limits & Tail Latency Collapse

4. The 2026 Unified vLLM-XLA Stack Architecture

Resources & Artifacts