Custom web apps and workflow automation, case studies from this year

A youth football league of 22 coaches, running on a Facebook group, a parent text chain, and three Google Sheets nobody owned. Registration stopped at signup. Everything between signup and game day was manual: equipment on paper, weigh-ins in somebody’s notebook, and no way to answer who still needs a weigh-in without calling around.

Thirteen modules across 80 phone-first screens: roster, draft, conditioning, practice planning, equipment, game day. React PWA with an offline outbox, so the sideline keeps working with no signal. Six composable roles, an audit record on every write. Claude Haiku turns a coach’s dictated notes into structured events, so nobody fights a form in week nine. That audit trail doubles as the usage record below.

Most software ships blind. We measured this one in week seven, against the live database.

The evening number killed the desktop-first admin view. We rebuilt for the sideline.

Live at tackleagent.com, running the league’s 2026 equipment day on a 248-player roster. Built pro bono, on the same standard as the paid work: 47 audited tables, 197 type-checked queries, a hard per-team cap on every Ai call. The flagship, because it’s the whole thesis in one system: build it, measure it, make it better.

A multi-troop program runs on a stack of unrelated SaaS subscriptions. One for registration, another for messaging, a third for trip planning, plus a Facebook group and three Google Sheets that nobody owns. Around $400 a month per program, four logins, and key data living in three places that drift out of sync.

Multi-tenant React + Express + Postgres. Opus for the reasoning-heavy work, Sonnet for the everyday flow, Haiku for high-volume classification: the right model per job, not loyalty to one. Per-tenant token caps. Backups verified nightly by restoring yesterday’s dump and diffing it against production, because an untested backup is not a backup.

Three separate subscriptions, four logins and a pile of spreadsheets, replaced by one system per program. Live at traillog.ai, in production with real Scout programs. Tri-model routing graduated out of Zero Drift and shipped here first.

A small shop runs more services than anyone can watch by hand. Every dependency (Claude API, n8n, Postgres, Docker images, dozens of CVE feeds) has its own retirement schedule, advisory list, and release cadence. Without a single screen, the failure mode is hearing about a problem from a customer instead of from a monitor.

FastAPI + React + Postgres. 34 modules sweep every dependency on each refresh: uptime, model retirements, CVEs cross-checked against our SBOM, token spend, and a backup-verify that restores yesterday’s dump to a throwaway database and diffs it against production. One Opus call writes the morning brief. We find out from a monitor, not from a customer.

Seventeen services across every system we run, watched on one screen. 99.97% availability over 90 days, and 127 deploys with zero rollbacks. The tile on the home page is the live view, not a screenshot.

A spec written by one author, with one model checking it, ships with contradictions and missing requirements that get expensive to fix once code is written. Single-model review is a single point of view. Single point of view is a single point of failure.

Eight phases from intake to traceability. Opus, GPT, Gemini, and Grok all run side-by-side on the same spec. We pick the one that’s right for the job, not the one we’re loyal to. The models debate, contradictions are surfaced, missing requirements are flagged. Full debate audit logs persisted alongside the spec so any decision can be traced back to which model said what and why.

Contradictions get caught before a line of code is written. Everything on this page came through Zero Drift first: TrailLog’s tri-model routing, Hopper Ops’ synthesis prompts, the pro bono builds’ intake flows. Nothing ships to a client until it has survived four models arguing about it.

Volunteer organizations serving hundreds of families hit the same wall: the coordination eats more hours than the program. A Scout troop with parent comms scattered across email, text, and a website nobody updates. Neither org has a budget for custom software, and both are exactly who never gets access to real engineering.

Real web apps, not portfolio exercises. Same scoping discipline, same documentation, same instrumentation, and the same monitoring standard as the paid work. Phone-first, because that is where the volunteers actually are. Built and donated start to finish.

TackleAgent shipped first and is the flagship case study above: live, instrumented, and measured against a real season. The Scout troop platform is in build on the same standard. Pro bono work gets the measurement layer too, which is how we know it is a standard and not a discount.