Blog

AI agents, RAG, and integrations — notes from production.

Vendor Lock-In in AI: Where It Hides and How to Limit It

AI vendor lock-in hides in prompts and flows in a vendor console, a fine-tuned model you can't export, conversation logs, and embeddings tied to one model.

How to Write an AI Requirements Doc Engineers Can Build From

An AI requirements document needs the task with examples, acceptance criteria as an eval-set spec, the non-functional constraints, and failure behaviour.

AI Incident Response: What to Do When the Model Is Wrong at Scale

AI incident response: detect from observability, triage the layer, contain with a feature flag to rules or humans, then root-cause and add a test case.

The Total Cost of Owning a Custom AI System

Total cost of ownership for AI is build plus run plus the lifecycle costs that outlive the build: retraining, monitoring, on-call, eval upkeep, model migration.

How AI Vendors Price Their Products — and How to Compare Them

AI vendor pricing comparison: per-seat, per-resolution, per-token, per-request. Normalise every quote to cost per unit of work at your volume, including peak.

Guardrails for an LLM in Production: Input, Output, and Action

LLM guardrails in production come in three layers: input (injection, PII, off-topic), output (schema, unsupported claims), and action (what the tools can do).

When to Fine-Tune vs When to Prompt Better

Fine-tune vs prompt engineering: fine-tuning buys format consistency and lower per-call cost, not new knowledge. Exhaust prompting, few-shot, and RAG first.

LLM Observability: What to Log and Alert On

LLM observability: trace each request end to end - prompt, retrieval, tool calls, output, tokens, latency - and alert on refusals, tool errors, cost, and drift.

Data Readiness: What Your Data Must Look Like Before an AI Project

Data readiness for an AI project: an API to reach the data, labelled outcomes, current content, and enough history - assess it before committing a timeline.

Measuring Whether an AI Feature Actually Moved the Business Metric

AI feature ROI measurement: use a holdout, define the metric and owner before launch, and net out deflection that isn't resolution, seasonality, re-contacts.