Blog

Engineering notes from Poket Dev. Python, AI/ML, cloud and DevOps, hand-written in Detroit.

Featured

Visual support for custom AI inference chips in an operational workflow.

Tools • Jun 24, 2026 •12 min read

Custom AI Inference Chips Are Eating the GPU Market

Discover why frontier AI labs are shifting to custom AI inference chips to solve memory bandwidth bottlenecks, reduce latency, and challenge GPU dominance.

Professional context for ai reasoning transparency.

Guides • Jun 22, 2026 •9 min read

The End of AI Reasoning Transparency: When Chain of Thought Becomes a Summary

AI reasoning transparency is failing. Learn why models like Claude summarize their chain of thought, why raw logic is hidden, and how to evaluate black box AI agents.

Conceptual visual related to AI agent data leakage.

News • Jun 22, 2026 •11 min read

Why AI Agents Leak Sensitive Data (and How to Stop Them)

Discover how autonomous AI agents leak sensitive enterprise data through reasoning traces. Learn practical architectures and sanitization frameworks to prevent exposure.

Recent

Prefill decode disaggregation isolates compute phases across separate GPU pools in a datacenter to optimize large language model serving.

Engineering • Aug 03, 2026 •11 min read

Disaggregated GPU Inference Hits the KV Cache Wall

Disaggregated GPU inference splits prefill and decode for higher throughput, but each request moves 2.6 GB of KV cache across the datacenter.

Document mediated prompt injection turns shared document stores into self-propagating attack surfaces when an AI agent both reads from and writes to the same files.

News • Aug 01, 2026 •11 min read

How The Copilot Prompt Injection Worm Spreads

The Copilot prompt injection worm proves prompt injection can self-propagate through shared documents. Learn why AI security fails and how to adapt.

LLM-native recommendation architecture using generative language models to replace hand-crafted feature pipelines in production recommendation systems.

Use Cases • Jul 31, 2026 •13 min read

LLM-Native Recommendation Architecture After Netflix GenRec

Netflix's GenRec replaces thousands of hand-crafted ML features with an LLM-native recommendation architecture. Explore the engineering trade-offs for builders.

MCP servers for AI agents provide a standard protocol that bridges fragmented operational data silos without per-system custom connectors.

Use Cases • Jul 30, 2026 •13 min read

How MCP Servers for AI Agents Bridge Fragmented Data

MCP servers for AI agents dissolve fragmented operational data silos, shifting the bottleneck from integration to query planning and context engineering.

Controlling reasoning effort in LLM APIs means setting explicit parameters that cap how many tokens a model spends on internal chain-of-thought before producing visible output.

Guides • Jul 28, 2026 •13 min read

Control Reasoning Effort LLM APIs in Production

Control reasoning effort LLM APIs across OpenAI, DeepSeek, and Anthropic. Practical routing rules to cut cost and latency without losing accuracy.

BM25, dense, and SPLADE retrieval baselines compared for production RAG, each with distinct tradeoffs in memory footprint, query latency, and failure behavior.

Engineering • Jul 27, 2026 •13 min read

BM25 vs Dense Retrieval and SPLADE for Production RAG

BM25 vs dense retrieval on a consumer GPU. Each baseline wins one RAG scenario and breaks another. Match your index to memory, latency, and failure mode.

AI-driven browser agent performing web automation tasks on a computer screen, representing the production challenges of shipping LLM browser agents.

Guides • Jul 26, 2026 •12 min read

Shipping LLM Browser Agents Without Breaking Production

LLM browser agents look easy in demos. Shipping one means defaulting to APIs, containing prompt injection, and budgeting for unpredictable per-step costs.

Claude Opus 5 achieves a 0% prompt injection success rate across 129 browser-agent scenarios, offering a model-attributed security metric for foundation model selection.

Tools • Jul 25, 2026 •10 min read

Claude Opus 5 Prompt Injection Hits 0% Across 129 Tests

Claude Opus 5 prompt injection tests show 0% success with Auto Mode versus 3.7% without across 129 scenarios, making resistance a model selection metric.

AI coding agents reduce per-task cost by indexing repository structure rather than re-exploring codebases from scratch on every session.

Tools • Jul 24, 2026 •12 min read

AI Coding Agents Cut Per-Task Cost Up to 4x With Indexing

Repo-aware AI coding agents cut per-task cost up to 4x by indexing repositories once, ending the cold-start tax that rebuilds context every session.

Rogue AI agent persistence enables malicious agents to operate indefinitely within trusted enterprise workspaces, evading traditional security controls.

News • Jul 23, 2026 •12 min read

ChatGPT Shared Link Vulnerability Plants Rogue Agents

The ChatGPT shared link vulnerability plants persistent rogue agents inside enterprise workspaces. Learn why agent persistence outlasts prompt injection.

Comparison of Grok and Copilot as AI assistants for building and automating spreadsheets in Excel.

Tools • Jul 22, 2026 •9 min read

Grok vs Copilot in Excel for Builders

Grok vs Copilot in Excel is a clash of integration architectures. Learn how Copilot's deep runtime compares to Grok's sidebar overlay for builders.

Apple Silicon Mac running large language model inference locally through the MLX framework and unified memory.

Guides • Jul 21, 2026 •13 min read

Run AI Models Locally on Mac With MLX and Nativ

Run AI models locally on Mac with MLX and Nativ. A trade-off framework for when on-device inference beats APIs on cost, privacy, and latency.

The Poket Dev Dispatch

Engineering notes from Detroit. Python, AI/ML, cloud and DevOps. No spam, unsubscribe anytime.

Join 3 readers. No spam. Unsubscribe anytime.