Blog
AI automation, vibe coding, SaaS, MVPs and browser automation, written from projects that shipped.
A valid webhook signature only proves who sent a request, not what that sender is allowed to change. This is the intake layer I build so webhooks can safely trigger automations and AI agents.
Read ArticleMost AI features are judged by demos and gut feeling. I track one thing instead: how much a human has to change the output before it ships. Here is how to capture it, compute it and use it to pick models and prompts.
Read ArticleMost bad RAG answers come from broken PDF extraction, not a weak model. Here is the ingestion layer I build: triage, coordinate-based extraction, structure rebuilding, per-page OCR fallback and quality gates.
Read ArticleLong plans drift and go stale. Failing acceptance tests give a coding agent a spec it can check itself against. Here is the workflow I use to steer agents without plan mode.
Read ArticleCoding agents forget your project every session. Here is the four-file system I keep in every repo so decisions, constraints and unfinished work survive between sessions and between models.
Read ArticleAgents that mutate your database directly are impossible to audit, debug or roll back. Here is the proposal ledger pattern I use: the agent writes intentions, a boring committer applies them.
Read ArticleTenant isolation is not enough once you add RAG. This is how I make AI search and assistants respect per-document permissions, revocations and derived data in multi-tenant SaaS.
Read ArticleMost agent systems have a Stop button that only stops the spinner. Here is how I build cancellation that actually halts loops, aborts provider calls, undoes side effects and caps spend.
Read ArticleA practical metering and cost-attribution setup for SaaS with LLM features: usage event schema, a money ledger, showback vs chargeback, and margin alarms that catch a loss-making customer before your invoice does.
Read ArticleMost agent failures I debug are wrong-tool failures, not reasoning failures. Here is how I design a tool surface: intent-based naming, coarse tools, schemas that block bad calls, progressive disclosure, and a tool-selection eval you can run in CI.
Read ArticleEach LLM call ships your customer's data to a third party. Here is the gateway, redaction map, policy table and test suite I use so that export is deliberate, logged and defensible.
Read ArticleMost AI MVPs die at the input, not the model. Here are six patterns I use to replace the blank prompt textarea with structured input that compiles into prompts - plus a two-day retrofit plan.
Read ArticleMost LLM classification work fails because the labels are undefined, not because the prompt is weak. Here is the workflow I use: build the taxonomy, extract facts instead of verdicts, and let code decide.
Read ArticleYour eval scored 94% and production is a mess. The gap is almost never the model - it is untracked drift between what you tested and what you deployed. Here is the manifest-based workflow I use to close it.
Read ArticleA running process tells you nothing about an AI pipeline. Here is the three-layer health check system I install in every automation: liveness, readiness, and capability canaries with golden inputs.
Read ArticleAgents can write more code than any team can review. Here is the concrete PR protocol I use - diff limits, blast-radius tiers, test-tamper detection, and a 12-minute human read that actually catches bugs.
Read ArticleMost scrapers have no concept of permission. I show the policy file, fetch gate, rate budgets and provenance logging I add to every parsing pipeline so it survives a complaint, a block or a client audit.
Read ArticleScrapers rarely crash. They quietly return plausible garbage. Here are the four validation layers I put in every parsing pipeline so bad data gets quarantined instead of shipped.
Read ArticleCoding agents write syntactically perfect migrations that lock your production table for four minutes. Here is the expand-contract workflow, the Postgres gotchas, and the rules file I feed the agent before it touches a schema.
Read ArticleAn AI agent is a long-running background job, not a web request. Here is the queue, run-state and checkpointing architecture I use in production, with schema, retry rules and cancellation.
Read ArticleIf your parser feeds scraped HTML into an LLM, you have an untrusted input problem. Here is the architecture I use to keep injected instructions from turning an extraction job into an action.
Read ArticleMost AI agents run as a human: shared logins, root API keys, the founder's browser profile. Here is how I give agents their own principal, scoped credentials, and a kill switch that actually works.
Read ArticleMost AI agents don't fail because of bad prompts - they fail because context grows unbounded and costs explode. Here's how I budget, compact, and store agent memory in real projects.
Read ArticleHow I run the same AI automation across a dozen clients without copy-pasting prompts or building one bloated mega-prompt: a four-layer composition model, per-tenant overlays with an explicit allowlist, golden sets, and version pinning.
Read ArticleMost broken AI automations aren't broken because the model is dumb - they're broken because every retry does the work twice. Here's the contract I use for idempotent pipelines, with schemas, hashing rules and an audit checklist.
Read ArticleCoding agents run shell commands, install packages and read your filesystem. Here is the exact sandbox, egress allowlist and dependency gate I use so a bad suggestion costs me a container, not my credentials.
Read ArticleMost AI products pay frontier prices for junk work. Here's the model router I put in front of every LLM call: routing tiers, a cheap classifier, caching, fallbacks, and per-feature cost accounting.
Read ArticleAI agents report success they never achieved. Here is the verification layer I put in front of every production agent - acceptance contracts, deterministic checks, evidence artifacts and retry budgets.
Read ArticlePrompts are not guardrails. Here's the enforcement layer I build in code - tool allowlists, idempotency keys, budgets, dry-run mode, approval gates and replayable traces - before an AI agent gets access to a client's production data.
Read ArticleFunctions are the containers that turn a pile of action blocks into a real program. How they work in BAS, why they take parameters and return results, and how they scale across threads.
Read ArticleModules are the boxes that hold everything BAS can do. Here is the LEGO mental model, the action-block → function → module hierarchy, and the core vs additional split.
Read ArticleBAS moves the cursor along a real trajectory shaped by three parameters - speed, gravity and deviation. How to tune them for raw performance or for passing behavioral bot detection.
Read ArticleThe mental model that makes Browser Automation Studio click: its four structural layers, and the two ways it talks to websites - when to render a browser and when to go raw HTTP.
Read ArticleCanvas fingerprinting tracks you even after you clear cookies. Here is why noise-based spoofing fails and how PerfectCanvas defeats it with real GPU renders.
Read ArticleA practical catalogue of what Browser Automation Studio actually builds - from auto-registration and parsers to checkers, mailers, monitors and desktop/Android automators.
Read ArticleA practical framework for turning a raw idea into a testable AI MVP in one month: scope cuts, architecture choices and the launch checklist I use on client projects.
Read ArticleWhat vibe coding really means in production work: where AI-driven development shines, where it fails, and the discipline that separates working products from demo videos.
Read ArticleAn honest guide to Browser Automation Studio: what BAS does well, where its limits are, and which business tasks it automates profitably.
Read ArticleNot every process deserves automation. Here is how to pick the first five that pay for themselves in weeks - and the three traps that burn budgets.
Read ArticleThe architecture decisions that let one person build, run and scale a SaaS product - and the fashionable choices that quietly kill solo projects.
Read ArticleGet a clear architecture, a working MVP and a system you can test, sell and scale.
I reply myself, usually within a few hours
Message me