I build systems.

● RESTRICTED · INTERNAL · UP SINCE 2026

I designed, built, and deployed it with help from AI coding tools, and supported it until I changed teams: a 26-container microservice platform on a spatial PostGIS core. Eighteen bounded services behind one gateway: project tracking, field operations, AI document extraction, client delivery. It replaced Asana for project and task management, and the department runs its field and office work on it. The map below is the real topology. A few were built and never put into daily use: the compliance gate, the analytics layer, and the field-to-office pipeline. Click around.

● CORE LOOP PLAYABLE · PERSONAL LAB · UNITY 6 + .NET · 1,500+ TESTS PASSING

Land Survey Simulator

A land survey simulator: the arithmetic is automated, the judgment is modeled.

Nobody teaches surveying to someone who has never held a prism pole, so I am building the thing that does. The brain is a plain .NET class library with no engine reference of any kind: survey mathematics, the error model, boundary rules, grading. It builds for net8.0 and netstandard2.1, has no package references at all, and the same assembly runs under dotnet test and inside Unity. More than 1,500 tests run in seconds, which is what makes it safe to keep changing the math. Coordinates are the part I got strict about early. Everything is stored as doubles in meters and radians in a projected CRS and converted only for display, and survey data never round-trips through a 32-bit float: at Texas state plane northings a float resolves to roughly half a foot, which is worse than the instrument being simulated. The engine transform is a rendering against a floating origin that re-bases as the player walks. Every surface the player sees is in US survey feet, which is exactly 1200/3937 meters and not the international foot. Two parts per million sounds like nothing until you are seven million feet north, where it is about fourteen feet. The fixture the whole suite is built on is a five station loop with analytically known geometry, enclosing 22,500 m² exactly. I computed every expected value independently before writing the library, because matching a published traverse proves you copied it correctly and proves nothing else.

The rule that constrains everything is that the player is never shown ground truth. Every observation is truth plus modeled error, and estimating an unknown from imperfect evidence is the whole game. The crew chief who reviews your work reasons only about the record you produced, never about the terrain, and a reflection test asserts that no truth-bearing type can reach his view of the job. That firewall is what keeps the feedback honest: he is allowed to say that you coded a top of bank and there is no toe anywhere in the job, and he is not allowed to say you missed the bank. Grading is measured rather than asserted. P95 deviation is the headline, not RMSE, because a missed breakline is a large error over a small area and RMSE averages it into invisibility. On the reference bank, leaving a three meter grade break unshot reads as 0.53 m RMSE against 1.38 m P95, and shooting top and toe takes both to zero.

The part I got wrong was the chief himself. Played on a correctly worked lot he produced nine remarks, most of them technically true and useless. A false positive costs more than a miss, because once you learn to dismiss him every real finding goes with the false ones. I narrowed the finders against the actual ground and measured his remarks back down from nine to two. Underneath the one job there is a career, written as a tested model rather than as shipped content: rodman, gun, party chief, licensed, your own firm, on the rule that a rung has to hand you a new way to work and never better numbers.

● PRE-LAUNCH · TWO BRANDS, ONE BACKEND · 114 TESTS, REAL FEDERAL DATA · contoursplease.com

A SaaS in the making: draw or upload a site boundary and one pass returns real USGS contour lines, a stamped CAD file in six export formats (DXF, KMZ, GeoJSON, Shapefile, SVG, PNG), a nine-section due-diligence report, and a first-person 3D walk of the actual terrain, aerial imagery draped over the lidar surface. Everything is computed from public federal data, never mocked: 3DEP elevation down to the one-meter lidar tier, FEMA flood hazard, SSURGO soils, NLCD canopy, TIGER roads, NHD hydrology. The algorithms are the point: marching-squares contour extraction, D8 drainage flow accumulation with ponding detection, a buildable-pad finder that hunts the flattest contiguous square under combined slope, water, and flood masks, a driveway grade profiler along any drawn line, and per-boundary NAD83 UTM zone selection by formula, correct anywhere in the continental US. The 114 backend tests run mostly against the real DEM, flood, and soils rasters, and the backend container is hardened: read-only rootfs, all capabilities dropped, non-root, tmpfs scratch. Two brands, one consumer and one professional, are served by a single backend and routed purely by nginx server name. Deliverables are planning-grade by design and say so on every page.

● RUNNING · PERSONAL LAB · C# / .NET · 227,060,311 PTS MAPPED

Autonomous characters living inside a real game server: traveling between towns, mining, crafting, trading on a persistent market ledger, buying houses, and holding LLM-backed conversations when spoken to. The map below shows agents routing themselves across a waypoint graph of the world, with their activity streaming beside it. And beneath the map: the same world as a 227-million-point reconstruction, rebuilt from the game's own data files and explorable in 3D (188M continent points shown).

● RUNNING · HOMELAB · 70+ CONTAINERS · 3M+ LOGS A WEEK

Real-time observability for every container on my home server: live log streams, health, uptime timelines, pattern alerts. The log hot-path is memory-only, a single background task batch-writes to SQLite each second, so bursts never stall the stream. Alerts fan out to dashboard, bell, and phone from one code path.

● RUNNING · HOMELAB · 2.5 GBPS MIRROR · 1.6M PPS TESTED

My ntopng replacement. A managed switch mirrors every packet in the house into a Go capture agent: AF_PACKET fanout workers, nDPI deep packet inspection, and device identity sniffed from DHCP, mDNS, DNS, and ARP. Flows pre-aggregate in the agent, stream over a unix socket (with a disk spool for backend downtime), and land in a single-writer SQLite pipeline behind the graph ntopng never gave me: every device and protocol at once, zoomable from live seconds to months. Replay benchmarks clear 1.6 million packets per second, four times the worst case a saturated mirror can produce.