Over the past few years, I've noticed AI-powered UAVs evolving from simple aerial cameras into intelligent systems that can inspect infrastructure, monitor crops, generate maps, detect anomalies, and even make decisions during autonomous flights.
The technology seems ready for much broader adoption, yet many industries still rely on traditional workflows.
From what I've seen, the potential benefits include:
Faster inspections with fewer safety risks
Better data collection for decision-making
Reduced operational costs through automation
More frequent monitoring of assets and infrastructure
Easy to kill a feature nobody wants. Much harder to kill one that would clearly get real usage, just not enough to be worth what it costs to build and support.
Went through this properly with a batch of automation features for letsflw. Started with 8 candidates, all genuinely useful on paper. Ended up shipping 3. The other 5 got cut for standalone demand that was real but thin the kind of thing where you'd get a trickle of users, permanently, and permanently maintain a feature for that trickle.
Yesterday, I came across this visualisation of skills required for AI-related positions on Instagram.
To be honest, it feels a bit outdated to me because AI can now handle most tasks quite well: writing texts, generating images, coding, and, to some extent, even creating videos.
Like most things, Loopify started out of frustration.
Cameron and I were just trying to find a simple way to post across multiple platforms, but what we found were either outdated tools that felt like flip-phone era relics, or shiny ones charging premium prices for the bare minimum. Or both. So we started building what we wished existed. Something modern, clean, and actually pleasant to use.
Right now, we re in learning mode, talking to creators, marketers, and small teams to understand what actually matters in a tool like this.
Analyzed a day of token logs across an autonomous coding agent setup running on internal codebases. The raw count: 769M input tokens vs 7.4M output tokens (~104:1).
Because long agent runs re-read session history (files, AST diffs, test outputs) every turn, input costs accounted for ~95% of total spend. Optimizing output length turns out to be looking at the wrong variable.
Three things actually saved us money:
Routing: Shifted non-interactive workloads (evals, background analysis) to batch/flex channels. Billed at 0.5x list price with zero code logic changes.
Cache Discipline: Kept system prompts strictly byte-stable (no top-level timestamps). Achieved a 94.9% prompt cache share, driving input costs from $10/M down to ~$1.46/M blended.
Context Compression: Built a pipeline sending compact session representations instead of verbatim transcripts. Achieved 2.83x median compression (fitting ~500K session history into a 200K window).
Many products add a setting as soon as two users ask for different behaviour. The result is often a page of switches that makes the common task harder. A better first step is to identify the most frequent, least surprising default and leave the edge case visible in the decision record. When have you chosen a strong default instead of an early setting, and what told you it was safe to do so?
For as long as software has existed, the user has been a person. Someone sitting at a desk, poking at a phone, or calling an API. That assumption was so obvious it was never really a design principle, it was just common sense. Every decision about hierarchy, color, button placement, and error messaging was downstream of a single fact: a human being is going to see this.
