Insights

Notes from the systems we're actually building.

Applied engineering writing, not marketing — LLMs, computer vision, agentic systems, generative AI, and the infrastructure underneath all of it.

NLP & Text Analytics6 min read

Stop Routing Every Text Task Through an LLM

Classifying support tickets, deduplicating leads, and tagging documents don't need a frontier model — and running them through one anyway is costing teams accuracy, not just money.

July 25, 2026

Document Intelligence7 min read

The OCR Step Isn't Optional, Even When You're Using an LLM

Vision-language models can read a document image directly, which makes it tempting to skip OCR entirely. In production, that shortcut is where document pipelines quietly start failing.

July 18, 2026

Voice & Conversational AI6 min read

Why Your Voice Agent Sounds Robotic (It's Not the Model)

Teams debugging a sluggish voice agent almost always look at the LLM first. The actual latency budget is usually being spent somewhere else entirely.

July 10, 2026

Geospatial AI7 min read

Why Your Satellite Imagery Model Works in the Notebook and Fails Over a New Region

A geospatial model that scores well in validation and then falls apart the moment it sees a different city isn't unlucky — it's the default outcome unless the pipeline is built to expect it.

June 29, 2026

Systems & Infrastructure7 min read

What Kernel Bypass Actually Buys You (And When It's Worth It)

eBPF and XDP get pitched as a universal performance upgrade. The honest version: kernel bypass is a serious engineering commitment that only pays off at a specific kind of scale.

June 15, 2026

LLM Engineering7 min read

Why Your RAG Pipeline Is Failing (And It's Not the Retrieval)

Most retrieval-augmented generation systems don't fail because the vector search is bad. They fail because nobody defined what a correct answer looks like.

June 2, 2026

Computer Vision6 min read

Computer Vision Models Die in Production. Here's Why.

A model that hits 94% mAP in the notebook and falls apart on the factory floor isn't a modeling failure — it's a data pipeline failure that happened months earlier.

May 14, 2026

Agentic Systems6 min read

MCP Is the Missing Layer Between Agents and Your Actual Systems

Every AI agent demo can call a tool. The hard part is doing that safely, with real permissions, against a system that has customers depending on it.

April 22, 2026

Generative AI6 min read

What We Learned Shipping a Generative Image Pipeline at Scale

Fine-tuning a diffusion model to match a brand is the easy part. Making it reliable enough to publish without a human checking every image is the actual product.

March 18, 2026

AI Infrastructure7 min read

The Real Cost of Running LLMs in Production (And How to Cut It)

Token pricing is the smallest line item in most LLM infrastructure bills. The real costs hide in retries, over-provisioned context, and eval you're not running.

February 9, 2026

LLM Engineering5 min read

Evaluation Is the Bottleneck Nobody Budgets For

Teams will spend months on prompt engineering and days on evaluation. That ratio is exactly backwards, and it's why so many LLM features stall in review.

January 12, 2026