Blog
Engineering deep-dives, opinion essays, and lessons from shipping AI for some of the world's most demanding companies. 25 articles and counting.
All articles
Retrieval is 80% of the answer quality and 5% of the demos. A field guide to chunking, reranking, citations, and the eight stages teams skip.
Sub-second latency isn't the goal. Predictable latency, with the right barge-in semantics, is what makes a voice agent feel human.
If you can't run a regression in CI, you don't have a feature — you have a demo. Building golden sets, judges, and drift detection.
Prompt injection is the new SQL injection. Seven layers we put between the user and your most sensitive systems.
The default answer to 'should we build it?' is no. The exceptions are where your moat lives.
You don't need ten PhDs. You need three senior engineers who have shipped one bad LLM product and learned why.
Cost-per-resolution, not 'productivity uplift'. A four-quadrant model that links AI work to a P&L line.
Frontier vs. open-weights vs. fine-tuned small. A pragmatic decision tree, with cost numbers from real workloads.
If the steps are known, write the steps. Agents are for the unknown — and they're more expensive than you think.
We won't translate the regulation. We'll tell you what to put in your system, in what order, and which auditor will ask for it.
Underwriting, KYC remediation, dispute triage, and one trap that every neobank falls into. Plus a benchmark of where the savings are.
Charting, coding, and triage all look like text problems until you read the case law. A practitioner's view on the real risks.
Token cost is the easy part. Idle compute, vector storage, and cold-start latency are where bills get out of hand.
If you've read three blog posts and still aren't sure what a vector is doing, this one is for you. With pictures.
STT, VAD, dialog manager, TTS — and the four boxes you have to add if you want it to work in a contact center.
Some of it survived. Most of it didn't. What actually moves model behavior, and what was just superstition.
Most HITL systems make humans into sad rubber stamps. Here's how to build review queues that humans actually want to use.
Fine-tuning is the expensive last resort, not the first lever. A checklist before you spend the GPU budget.
How to triage 70 ideas down to 6 funded projects in 8 weeks, without crushing the people who proposed them.
You don't need a lakehouse. You need three datasets cleaned, labeled, and exportable. Here's the audit we run.
An anonymized walkthrough of a deployment that took support contact rates from 18% to 7%, with the gotchas.
Your APM doesn't know what a hallucination is. What to log, what to sample, and how to find a regression at 3 a.m.
Frontier models get the headlines; 7B fine-tunes do most of the work. Where small models are quietly winning in 2026.
Twenty-three questions that separate vendors from poseurs. Print this, take it into your next meeting.
Talk to our team about a 2-week assessment.