Articles by Daniel
10 published
AI · Sep 18, 2026 · 5 min
How we shipped a support agent that resolves 62% of tickets
The architecture, evals and guardrails behind a production AI agent — and the three mistakes we made first.
AI · Jul 28, 2026 · 5 min
RAG for small teams: a knowledge base in two weeks
A pragmatic stack for turning your docs into an assistant your team actually trusts.
Engineering · Jun 12, 2026 · 5 min
Idempotent webhooks: building integrations that tolerate retries
Webhooks arrive late, out of order and more than once. How we design handlers and integrations that stay correct whatever the sender does.
AI · May 12, 2026 · 5 min
Fine-tuning versus better prompting: a practical decision guide
When a fine-tuned model is worth the data work and maintenance, and when prompting, retrieval and a good eval set get you there faster.
AI · Feb 24, 2026 · 5 min
Building an AI sales assistant that reps keep using
Why most sales copilots are abandoned within a month, and the design choices that made one stick with a 40-person sales team.
AI · Dec 09, 2025 · 5 min
Voice agents on the phone line: latency, handoffs and honesty
What it takes to put an AI voice agent on a real phone line, from sub-second response times to clean transfers and clear disclosure.
AI · Oct 15, 2025 · 5 min
Giving AI agents tool access without handing over the keys
Scoped credentials, server-side policy, confirmation steps and prompt-injection defenses for agents that can take real actions in your systems.
AI · May 27, 2025 · 5 min
When workflow automation beats an AI agent, and when it does not
A decision framework for choosing between deterministic automation, a single LLM step and a full agent, with cost and reliability trade-offs.
AI · Mar 19, 2025 · 5 min
Automating invoice intake with LLMs without losing the audit trail
A document-processing pipeline that extracts invoice data reliably, flags what it is unsure about, and keeps finance and auditors comfortable.
AI · Jan 14, 2025 · 5 min
Write the evals before the prompt: testing LLM features properly
How we build evaluation sets, graders and CI gates for LLM features so that every prompt or model change is measured instead of guessed.
