I killed my SaaS yesterday after 5 weeks and $0 revenue. Today I shipped its replacement. Start to finish in one day — demand research, build, adversarial security testing, payments, launch. The difference: this time I found the demand before writing code. 🧵

Dev log: an AI agent designs a road on real USGS terrain through MCP tools. Every frame in the video comes from the actual run. Route alternatives, alignment with spirals, profile optimised for cut/fill balance, pavement design, retaining walls, slope stability, culverts, storm + sanitary sewers, signals, a 38-sheet drawing set. Standards profiles for the US, Canada, UK, Australia and New Zealand, each check cited to its clause. The agent can't approve its own work. Only an engineer can, and only with zero mandatory failures.

Dev log #3: ulog-why. It reads a PX4 flight log and reports the earliest likely cause, with the evidence chain up to the failsafe. Test: PX4 v1.15.4 SITL flights, one injected fault each (GPS loss, baro/mag offset, thrust loss, battery failure, datalink loss). 50 held-out flights, scored once after freezing the rules: 45 right. Median 0.1 s from fault to detection. 0 false alarms on 7 clean flights. All simulation. Real crashes will be messier, and takeoff transients caused 2 of the 5 misses.

Dev log #2: mcp-lint. One Rust binary reads your MCP configs (Claude Code, Claude Desktop, Cursor, VS Code, Windsurf), starts each stdio server, and flags tool descriptions that give your agent orders. It also pins a SHA-256 of every tool, so a description that changes later shows up in `mcp-lint diff`. On 122 tools from 8 popular MCP servers: 0 default-level flags. On a test set I wrote after freezing the rules: it caught only 5 of 15 malicious descriptions. So treat it as a cheap first filter and change alarm. Bawbel (Python) did this first and covers more attack types.

Dev log: Claude Code's compaction can keep the numbers and drop the reasoning, open tasks and subagent results (claude-code issues #102666, #32099, #31420). So I added a "compact fuse" to ContextSpindle. A PreCompact hook saves the goal, open tasks, constraints, decisions and subagent results to a local ledger. A SessionStart hook puts them back after compaction. On a synthetic 185 KB, 95-message transcript: 14/14 test facts came back in 1.7 KB of context. Save 2.5 ms, restore 2.1 ms. Limits: keyword extraction (no model), so reasoning without trigger words gets missed. Not yet tried in a live Claude Code session.

Dev log: I gave a drone a fruit fly's escape reflex and tested it in PX4 SITL. Event camera → time-to-contact stage → giant-fiber model from the FlyWire + MANC connectomes → 1.2 s vertical climb. v4, 66 held-out trials per speed: 6–30 m/s: 0 collisions in each tier 60 / 90 / 120 m/s: 7.6% / 4.5% / 21.2% False triggers: 0/198 So the validated envelope is ≤30 m/s. Simulation only for now.