Writing
Notes from the build.
Our earlier posts covered blockchain topics and have been retired — they no longer reflect what we do. New writing on retrieval quality, evaluation and automation reliability is in progress.
What is coming
Practitioner writing on the parts of AI engineering that are genuinely hard, with real numbers where we are allowed to publish them.
Planned pieces include: why chunking strategy matters more than model choice in most RAG systems, what an evaluation harness actually needs to contain, the real cost curve of a voice agent at volume, and an honest account of what AI search visibility can and cannot be measured.
We would rather publish four posts a year that are worth reading than forty that restate vendor documentation.
Chunking beats model choice in most RAG systems
If a retrieval system is answering badly, the cause is usually that the right passage was never retrieved — swapping the model cannot fix a retrieval miss.
Fixed-size chunking splits a table from its header and a clause from its defined terms, and no amount of model capability recovers information that was never in the context window. Structure-aware chunking, overlap tuned to the document type, and metadata filters that constrain the search space usually move answer quality further than a model upgrade costing ten times more per token.
The diagnostic is cheap: take fifty questions the system got wrong and check whether the correct passage appears anywhere in the retrieved set. If it does not, the problem is retrieval. If it does and the answer is still wrong, then look at the model. Teams skip this check constantly. More on RAG and LLM integration.
What an evaluation harness actually has to contain
A test set drawn from real user questions, a scoring method a human agrees with, and a regression gate that runs before every deploy.
Most teams have a spreadsheet of questions someone wrote in an afternoon. That is not an evaluation harness, because it does not reflect the distribution of things real users ask, and it is never run again after launch. A usable harness needs cases mined from actual logs or support tickets, deliberately including the awkward ones — ambiguous phrasing, out-of-scope requests, questions with no correct answer.
It also needs a scoring method you have validated against human judgement on a sample. LLM-as-judge is fine, but only once you have checked it agrees with your own reviewers on the cases where the two might reasonably differ. Then it becomes a gate: if a change drops the score, the change does not ship.
The real cost curve of a voice agent at volume
Per-minute model cost is rarely the dominant line — telephony, retries, and the human escalation path usually are.
Pilots are priced on the demo call. Production is priced on the calls that go wrong: unanswered numbers that retry, callers who hang up at nine seconds, conversations that escalate to a human and therefore cost both the automated and the human minute. Modelling only the happy path understates real cost substantially, and it is the reason voice pilots so often look great and scale badly.
Latency is the other half. Every hundred milliseconds added between the caller finishing and the agent starting is felt as hesitation, and hesitation drives hang-ups. That budget constrains architecture — it decides how much retrieval you can afford mid-call and what has to be precomputed. More on voice agents.
What AI search visibility can and cannot be measured
You can measure whether AI assistants cite you and what they say; you cannot get reliable ranking positions, because there is no ranking.
Assistant answers vary by phrasing, by user context and by run. Anyone selling you a rank tracker for AI answers is selling a number that does not exist. What is measurable is citation share across a fixed prompt set sampled repeatedly, factual accuracy of what is being said about you, and referral traffic from assistant surfaces where the platform passes it.
The work that moves those measures is unglamorous and mostly overlaps with good technical SEO: clean structured data, content that answers a question in its first two sentences, and factual consistency across every page that mentions the same claim. See AI search visibility and SEO.
In the meantime, talk to us directly
Tell us what you are trying to build. We will tell you honestly whether we are the right team for it, and what it would realistically take.