Blog
Field notes on shipping AI-native systems with human engineering judgement: architecture, modernisation, automation, and the tradeoffs in between.
Most construction software content is a vendor listicle. This is the buyer’s and builder’s view: the AEC stack layer by layer, and the field-reality layer where off-the-shelf tools stop and custom work starts.
The demo flies on a toy repo. The 200k-line system nobody wants to touch is where AI coding tools stall. This is a practitioner’s guide to which tool class survives contact with production you can’t afford to break, the failure modes vendors won’t name, and a decision framework that matches tool autonomy to codebase risk.
You book the primary sale and still barely know what moved off the shelf. Secondary sales automation closes that gap: the SFA and DMS layer, the pipeline that survives distributor file chaos, and the reconciliation that makes the number trustworthy. Here is the whole system, mapped from our client work.
The build vs buy decision for secondary sales automation is rarely lost on the ingestion pipeline everyone worries about. It is lost on adoption in the field, and on the maintenance bill nobody priced in. Here is the framework we use in client work to make the call.
Every distributor sends secondary sales data in a slightly different shape, and the pipeline you build to read it spends most of its life absorbing that variance. Here is how we treat file-format normalization as the real engineering problem, and the failure modes we have watched break in production.
Primary sales are clean because they are billed from your own systems. Secondary sales are the number that tells you what the market actually bought, and they are the hardest thing in distribution to capture. Here is why the last mile breaks first, and what we have learned rebuilding these pipelines in client work.
How pharma secondary-sales pipelines fail - ingestion, parsing, validation, reconciliation - and the controls that fix them.
Break down voice-agent latency across VAD, STT, LLM, TTS and network; prioritize p95, fix endpointing and lower LLM TTFT.
Agentic IDEs can break legacy refactors-do scope limits, dependency mapping, characterization tests, and rules first.
Pick the smallest model and simplest runtime to self-host a private coding assistant-match model size, VRAM, and security to your use case.
A 2026 guide comparing Figma-to-code tools for handoff, prototypes, repo-mapped production - engineering cleanup remains necessary.
Shows how repeated inputs and parallel agents drive token costs-and how caching, scoping, and model routing cut spend.