Regnard Raquedan

TechGuides is the outside-of-work project closest to home for me. My wife Liza and I run it for families and seniors in Don Mills and Thorncliffe Park in Toronto. We relaunched it in April 2026 with fresh AI-focused content, a live site, and a learning tool we built ourselves.

The thinking is simple: The people with the most to gain from AI right now — families navigating a new country, seniors trying not to get left behind — are the people least likely to be served by the tools being built for them. TechGuides is our attempt to fix a small piece of that, in one neighborhood, with our own hands. If it works here, it can work elsewhere.

Precomputed AI, or PAI, is a design pattern I've been refining across the tools I build: relocate LLM reasoning into artifacts produced ahead of time, and reserve live inference for opt-in escalation. Reason ahead of time, serve instantly.

The pattern started as a footnote in a post I wrote about token consumption anxiety. The more I shipped, the more I noticed the same posture underneath each tool — precompute the common case, escalate only when the artifact can't decide. CloudEstimate, RightModel, and RunWhere are the worked examples. PAI is the name for what they have in common, and the pattern I'll keep writing and building against.

CloudEstimate is an open source tool I built because sizing self-managed enterprise workloads across clouds is harder than it should be. Vendor docs give you reference architectures. Cloud calculators give you pricing. Very little bridges the two quickly when the real question is whether the same workload looks different on Google Cloud, AWS, and Azure.

I spend enough time in that gap to know how much friction it creates. So I built something intentionally narrow: a way to map published guidance to concrete instance shapes and pricing snapshots, so a first-pass estimate can stand up in a Slack thread, a procurement review, or an architecture conversation. It's also a worked example of Precomputed AI — scheduled regeneration of the pricing artifact, served instantly at request time.

RightModel is a free, open source tool I built because the model-selection problem got noisy fast. Teams are being asked to choose between Anthropic, Google, OpenAI, and everything else before they even have a clear way to describe the work. Too often, that decision gets made with instinct, stale pricing tables, or whichever model someone happened to use last week.

So I built something intentionally simple: paste what you are about to build, choose whether you're constrained to a provider, and RightModel recommends a model in seconds with transparent reasoning, cost breakdowns, and a deeper path when the answer is less obvious. It's another worked example of Precomputed AI: a precomputed ruleset handles the common case in zero tokens, and a live LLM is there as opt-in escalation when the ruleset can't decide.

RunWhere is a free, open source tool I built for the AI infrastructure question that tends to turn into a spreadsheet too early: should this workload stay on a hosted API, or is it worth running your own model? Its default answer is intentionally opinionated: stay on the API unless your workload is one of the exceptions.

The site asks four questions — API spend, hosted model, traffic shape, and hard constraints — then checks whether the workload belongs in the exception set. When it does, RunWhere compares the relevant paths: managed endpoints, serverless GPU, always-on GPU VMs, scheduled batch GPU, and owned hardware. It's the third worked example of Precomputed AI: bake the boundary ahead of time, serve the default instantly, and reserve live analysis for close calls.