The notebook

All AI & LLMs Cloud & DevOps Industry Insights Laravel & PHP Modernisation Performance & Scale SaaS & Product Security & Compliance

AI & LLMs 20 posts

22 Jun 2026 AI & LLMs Per-User Token Budgets: Controlling LLM Costs in Your SaaS Provider dashboards tell you what you spent, not who spent it. How I meter tokens per tenant in Laravel with middleware, soft and hard quotas, budget alerts and graceful degradation. → 15 Jun 2026 AI & LLMs Cutting LLM Token Costs in Production: A 2026 Field Guide The five levers I apply to cut LLM API bills in production — prompt caching, model routing, batching, context compaction and output caps — ranked by effort-to-savings. → 01 May 2026 AI & LLMs Reliable JSON from LLMs: Structured Output Patterns That Hold Up The technique post underpinning every document-AI and integration project: native structured-output/JSON-schema modes vs the tool-call trick, validate-and-retry loops, enum constraints to ki... → 28 Apr 2026 AI & LLMs Invoice Data Extraction with LLMs: Beyond OCR in 2026 Why vision LLMs beat template-based OCR (reading text vs understanding which number is the total), a production pattern of JSON-schema extraction plus confidence-based human review, and the... → 25 Apr 2026 AI & LLMs Adding AI Features to an Existing SaaS: An Architecture Checklist The retrofit checklist I run on Laravel codebases: queue-based LLM calls (never inline in requests), timeout and fallback paths for provider outages, per-tenant cost ceilings, PII scrubbing... → 19 Apr 2026 AI & LLMs LLM Guardrails in Production: Surviving Drift and Bad Outputs Runtime reliability engineering as distinct from pre-launch evals: schema validation with retry loops, input sanitisation against prompt injection, fallback chains when a provider degrades,... → 16 Apr 2026 AI & LLMs LLM Evals Without an ML Team: A Practical Setup for Small Teams How a two-person team ships tested AI features: a golden dataset of 200-500 real (not synthetic) examples, assertion-style checks plus LLM-as-judge with its known biases, wired into CI the s... → 13 Apr 2026 AI & LLMs Human-in-the-Loop AI Agents: Approval Gates That Actually Work Where to place approval steps in agent workflows (outbound emails, invoice postings, CRM writes) using confidence thresholds, action whitelists, and audit trails — so automation saves time w... → 10 Apr 2026 AI & LLMs AI Agents vs Workflow Automation: A Decision Guide for SMBs A framework for when a deterministic queue/cron pipeline beats an agent and vice versa — variability, judgment, and natural-language inputs are the agent signals — with honest failure-mode a... → 07 Apr 2026 AI & LLMs RAG vs Long Context: When You Don't Need a Vector Database With million-token context windows and prompt caching, corpora under a few hundred pages are often cheaper and more accurate stuffed into context than retrieved — I give the cost/latency cro... → 04 Apr 2026 AI & LLMs Why RAG Fails in Production (and How to Fix Retrieval First) Roughly 73% of RAG failures are retrieval, not generation — a debugging checklist from client projects: inspect retrieved chunks before blaming the model, fix chunking (semantic over fixed-s... → 01 Apr 2026 AI & LLMs Small Models First: Why Cheap LLMs Handle 80% of SaaS AI Features An escalation architecture I use: default every request to a Haiku/Flash-class model at cents per million tokens, escalate to a frontier model only on validation failure or low confidence —... →