Direction
Running AI like a team
Clear briefs, bounded tasks, parallel work, and no output trusted until it is verified.
Seen in a per-discipline generation swarm and a repeatable method for writing agent briefs.
A Builder's White-Paper · Dec 2025 to Aug 2026
From Prompt to Platform
All of it built and run by one engineer: an education platform that writes its own curricula, the paid product on top of it, a live multiplayer game, client sites, and the infrastructure and operator system underneath. This paper is the full scope: the stack, the certifications, the judgment, and the mistakes that became doctrine.
8moDec 2025 to Aug 2026
13repositories, one account
2,218commits across the portfolio
749merged pull requests
6domains live in production
5+languages, Python to PowerShell
Certified by Anthropic
Three Claude certifications plus the Partner Badge, earned across the Architect and Developer tracks. The practice they certify is the one this paper documents.
Claude Partner Badge: Claude Code
Claude Certified Architect: Foundations
Claude Certified Architect: Professional
Claude Certified Developer: Foundations
01 · The Arc
The story these repositories tell is range: client web products, a real-time multiplayer game, an education engine with a commercial product on it, automation libraries from years of MSP work, and the infrastructure and operator tooling that keep all of it running. Every figure below is pulled live from the GitHub API, and the ledger in section six carries the exact counts.
By spring the ambition consolidated into a stack. I started the chatbot framework, then the education engine on top of it, then the commercial fitness module inside that, then the operator system I use to build all of it. In May and June alone I founded six new repositories, because by then I was building platforms instead of pages.
A second thread ran in parallel. The Anthropic path started with the AI Fluency course and its 4D framework: Delegation, Description, Discernment, and Diligence. Those four became the working discipline behind the build, and the road they opened ran through Anthropic Academy to Claude Certified Architect and a Partner badge.
2025 · run-up
capaz
Dec 2025
benesseremedestetica · wellness web
Cerca · closer.quest
Apr 2026
EGI-Codex · automation catalog
May 2026
EGI_bot · workflowtech.ai
GestaltWorkframe · the framework
GestaltWorkframeEDU · the platform
Jun 2026
llm-builder-kit · operator system
4vida · commercial module
msp-automation-api · API-ready MSP library
Jul 2026
ROFLcopter · roflcopter.cloud
iambraun.com · CV overhaul
Aug 2026
claude-cert-examprep · cert-prep coach skill
GestaltWorkframeEDU · commits per month
the depth ramp
866 commits and 468 merged pull requests on the flagship. *August measured through the 3rd.
02 · The Craft
Five technical capability areas, each stated as a claim and backed by what actually shipped. The judgment that directs them comes next.
Capability
LLM-native systems, grounded in real sources
Generation is the product here, and it is held to a standard of evidence.
- Built a self-generating curriculum engine that takes a subject, infers the pedagogy and domain shape, and produces graded programs with assessments.
- Grounded every generation in a real citation corpus, proven on production: 28MB of real book text across 80 sources, cited, with zero fabricated identifiers.
- Replaced a single generalist generator with a concurrent expert swarm, one corpus-scoped specialist per discipline: strength, mobility, conditioning, self-defense, yoga, breath, recovery.
- Made background generation survive deploys, re-dispatching orphaned runs on restart instead of leaving them stuck.
Verification standardA generation that reports success can still be wrong. So no program counts as finished until it has been walked through its real endpoints. That walk caught crashes a 96 percent progress bar never would.
Capability
Full-stack platform engineering
The whole stack ships from one desk, database to browser, and the platform boundary is part of the design: an engine that builds programs, and programs that snap into it cleanly.
- FastAPI, SQLAlchemy 2.0 async, SQLite, and Alembic behind a Next.js 15 static export with Tailwind v4 and TypeScript.
- Authentication as passwordless email-code sign-in, with roles enforced server-side and admin signup gated to the platform host.
- A domain-agnostic engine whose content modules carry all domain configuration and no engine code: the paid fitness product ships as pure config, and CI proves the engine domain-agnostic by running an English-course module through it unchanged.
- A module toolchain (.gwmodule bundles, a modulectl CLI, a framework-compatibility gate, atomic installs) so a finished program installs into a host platform in one verified step.
ReachThe same stack recurs across the account: the learning platform, the wellness site, the relationship app, and the capability tracker.
Capability
Commercial & monetization engineering
The money layer is real: live payment integrations and unit economics that get measured.
- Freemium architecture: a free multi-week teaser, the full program behind a paywall.
- An entitlement spine: one record per user and module tracking status, source, and period end, and every access gate is built to read it.
- Stripe checkout across four billing durations with webhook reconciliation, plus PayPal with real webhook signature verification.
- A regeneration credit ledger and reuse-first economics: catalog reuse makes zero extra model calls, cutting the cost of a full program generation from $3.40 to under a dollar.
Capability
Infrastructure & DevOps
The products run on infrastructure of my own, and keeping it running is part of the work.
- A private WireGuard mesh linking the workstation, two IONOS VPS hosts, and a NAS.
- Self-hosted GitHub Actions runners per private repo, parallel pytest with xdist, path-filtered workflows, concurrency auto-cancel.
- systemd services, nginx with per-endpoint routing and certbot TLS, Docker images, and a manual gated production deploy.
- Operations discipline from production incidents: reconciling SQLite schema drift on boot, tuning VPS swap after an out-of-memory hang, and diagnosing a runner that reads offline while its service shows active.
Capability
The operator system, and the process it enforces
The system used to build everything else is itself one of the builds. 700+ merged pull requests have passed through it, each through the same gate: dispatcher-assigned to a coding agent, reviewed, then merged by hand.
- llm-builder-kit: my doctrine plus the machine that enforces it. A dispatcher hands bounded tasks to coding agents, and a reviewer chain signs off before anything merges.
- The same kit runs a token-gated Next.js panel with background workers for routing, spend, and daily briefs, an Avalonia desktop cockpit over all of it, a self-hosted LiteLLM key broker, and standards vendored into every managed repo so drift fails CI.
- A model-routing policy that picks the cheapest tier that can do the job and escalates only with a stated reason, expected value, budget, and fallback.
- The gate itself: branch, PR, CI green, squash-merge. Clean history as a standing rule.
- Polyglot delivery: Python and TypeScript at the core, with PowerShell, Go, JavaScript, and Shell across the portfolio, including an API-ready MSP automation library (idempotent modules with rollbacks, Prometheus metrics, OpenAPI contracts, Pester tests) built from years of MSP work.
03 · Case Study: Engineered to Stay Up
ROFLcopter (a real-time multiplayer party word game, live in production) stays up because keeping it up was engineered in from the start. First the product, then the system underneath that keeps it standing.
The product
A humor game that runs itself
Players race a set of random letters into the funniest phrase and vote on the winner. Empty seats fill with LLM bots, each writing in its own comedic voice; a panel of independent AI judges scores every round and learns what a crowd finds funny from real votes, storing no identities. What the judges learn feeds the writers too, and the whole game runs in English and Spanish, bots and judges included.
- It never simply breaks. If the model key is missing, a call fails, or the batch has not warmed, the bots fall back to an offline part-of-speech generator and the game keeps moving.
- Deterministic by construction. A framework-free engine takes an injectable clock: production uses the real one; tests use a virtual clock that runs a full eight-round game plus the finale instantly, so even the finale is verified end to end with no device and no wall-clock wait.
- Drop-proof rooms. A bot instantly takes the seat of anyone who leaves mid-game, so the room and the finale's jury never thin out because a player bailed, and quitting carries a penalty.
- Defensive parsing, taught by production. Models violate their own structured-output contracts. One wrapped its JSON in a markdown fence and silently dumped whole batches to the fallback, until the parser learned to unwrap fences, accept stray shapes, and line-split as a last resort.
The system that keeps it up
The measured capacity ceiling
One box behind a single event loop is cheap to run and has an honest ceiling. The whole discipline is refusing to meet that ceiling for the first time during an outage.
- The box reports on itself. Event-loop lag, the binding constraint of a single-process real-time server, is sampled continuously and charted against its ceiling on a token-gated owner panel, next to model-call latency, live rooms, and multi-week player and funnel trends.
- The ceiling has a number. A standalone headless harness ramps simulated rooms against production to measure the real capacity ceiling instead of guessing at it.
- Config, not code. The moderation floor, the model identifiers, and the API keys are server configuration. Retuning the safety list or swapping a model is an operations task; no release required.
- Boring deploys. A merge to main builds and ships on a self-hosted runner; a docs-only change deploys without restarting the live service.
04 · Case Study: Observed, Measured, Recorded
The ROFLcopter study is about staying up. This one is about choosing and paying correctly. The system is a production inference layer built and operated at EGI, routing four workload shapes (long-form documents, interactive chat, short lookups, and tool calls) across a live catalogue of 360+ models. The full design is its own white-paper; what belongs here are the decisions that transfer, and the numbers that earned them.
The design
No model name in application code
A pinned model name is a stale constant the day something better ships. So a request declares what it needs and what it is optimizing for, and the layer resolves both against the live catalogue at request time. When something better ships, it wins on the next refresh with no code change.
- Requirements are machine-checkable: tool-calling support, a context floor, an output ceiling that will not truncate, and quality floors read from published benchmark indices. One catalogue entry advertises optimization for tool use and measures 18.9 on the agentic index against a median of 20.8; a floor catches what a description sells.
- Objectives stay separate from requirements. Nearly every early routing defect was those two axes collapsed into one control: maximum capability on a one-line lookup priced it like a full document, and fastest-option on a document handed a twelve-thousand-token job to a model that stops at four thousand. Same bug, two faces.
- Every decision is recorded. Per workload, a health payload carries the model chosen, its margin over the requirement, cost per turn, estimated seconds, and which tie-break decided. Automatic selection without a visible decision record cannot be debugged and cannot be defended.
The measurements
What the instruments found
Published operating data failed on contact: every latency field was null on all 16 endpoints checked, and one model reported 100 percent uptime while failing calls. So the layer trusts only its own instruments, and what they found paid for them.
- An active speed probe broke a circular trap. An unmeasured model never wins a speed-sensitive objective, so it is never selected, so it is never measured. The probe timed the fastest candidate at 658 to 833 ms to first token; a third candidate measured 11.6 seconds and had been winning a lane on price alone. Unusable for an interactive lookup, and nothing published would have said so.
- The objective, not the model list, sets the bill. The same short lookup costs about $0.047 at maximum capability against about $0.001 at soonest-per-dollar, a spread of roughly fifty times with the same requirements satisfied. Measured across all four workloads: $0.0009 to $0.374 per turn, and the full measurement sweep itself costs about $0.0003.
- Delivery time turned out to be part of the price. 28 catalogue entries are asynchronous batch endpoints: identical model, identical benchmark scores, half the price, delivered in hours. On any cost-ranked objective one wins permanently, and no quality floor can catch it. Arrival time was the defect. The price was fine.
- When the data said so, I deleted my own feature. Roughly 400 lines of discount-detection machinery, built and tested, came out after one check: aggregator prices already arrive net of discounts, so the machinery had nothing to detect. The finished router is smaller than the one it replaced, which is the version a client wants to inherit.
Why this transfersThe application behind this layer is a tabletop tool. The routing, the observability, and the cost control are the parts that move to any system that pays for inference by the token.
05 · The Judgment
The repositories show what I can build. This is how I decide what is worth building, and how I know when it is actually done. It is applied AI fluency: the four habits Anthropic names Delegation, Description, Discernment, and Diligence, learned the same way the technical skills were: by doing the work.
Taste
Knowing what to build, and what to leave alone
Product judgment is deciding which problems are real: telling correct behavior from a defect, and a business decision from a bug.
Seen in reading a "skipped" warm-up as correct drill placement, closing a list of audit defects that were mostly phantom, and treating a missing figure as a copyright call rather than a code fix.
Honesty
Calibrated confidence
Verified and assumed are different categories of claim, and the difference gets stated out loud. The most expensive failure in this work is a confident claim built on an unchecked assumption.
Seen in a memory system that records corrections so the same mistake is not made twice, and a shipped white-paper that offers to go through the parts that are still wrong.
Economics
Spending like it is my own money
Every model call and every server hour is a real cost. Work routes to the cheapest tier that can do the job and escalates only with a stated reason, an expected value, and a fallback.
Seen in a written model-routing policy, and a reuse-first design that made regenerating from scratch the rare path instead of the default.
Ownership
Deciding, and carrying the call through
Working solo, the call is mine: make it, document it, take it all the way to production.
Seen in an operator doctrine that draws a clear line between what ships autonomously and the few irreversible calls that earn a second look.
Resilience
Steady under real operations
Production breaks. The discipline is diagnosis before touching anything, load paced so systems stay up, and recovery without panic.
Seen in recovering a server from an out-of-memory hang, adding swap, and the rule to never stack a heavy generation, CI, and a deploy at the same time.
Communication
Translating across audiences and languages
The same work has to read one way to the engine, another to the buyer, and another to the person using it. I write the spec, the white-paper, and the product copy, and I build for English, Spanish, and Italian users.
Seen in a consolidated platform spec, a monetization architecture doc, published product briefs, and a personal writing voice I defined and keep.
Learning
Teaching myself, fast
The through-line under all of it. From a first app to production platforms in months, while building an engine whose whole job is how people learn.
Seen in the arc in section one, and a pedagogy layer that works out how a subject should be taught before it teaches it.
06 · The Portfolio Ledger
Every figure below is pulled from the GitHub API. Commits and merged pull requests are exact.
Core stack: the EGI platform
| Repository | Language | Commits | PRs | Status |
|---|---|---|---|---|
GestaltWorkframeEDU Self-hosted modular LMS that generates source-cited curricula and publishes them | Python · TS | 866 | 468 | Engine |
EGI_bot The workflowtech.ai agent: capability-based model routing over OpenRouter, cited automation CODEX | Python | 454 | 69 | Live |
llm-builder-kit Agent factory, desktop cockpit, key broker, standards vendored across repos | TypeScript | 278 | 13 | Live |
GestaltWorkframe Upstream chatbot framework: branded bundles, connector retrieval, budget-capped routing | Python · TS | 57 | 32 | Active |
Products & client work
| Repository | Language | Commits | PRs | Status |
|---|---|---|---|---|
ROFLcopter Bilingual real-time party word game with LLM comedy bots, at roflcopter.cloud | TypeScript | 151 | 97 | Live |
4vida Paid bilingual training and nutrition product, pure config on the EDU engine | YAML · MD | 97 | 38 | Live |
benesseremedestetica Wellness & aesthetic-medicine web platform | TS · MDX | 72 | 0 | Live |
capaz Capability tracker | TypeScript | 39 | 0 | In dev |
Cerca Bilingual relationship app on the "36 Questions" research, in beta at closer.quest (repo: closer) | TypeScript | 12 | 0 | Beta |
Libraries & tools
| Repository | Language | Commits | PRs | Status |
|---|---|---|---|---|
msp-automation-api API-ready PowerShell automation for MSP orchestration platforms: idempotent modules with rollbacks, metrics, schema contracts, Pester tests | PowerShell | 5 | 0 | Public |
claude-cert-examprep Exam-prep coach skill for Anthropic's Claude certifications (CCAR-F, CCAR-P, CCDV-F), three lanes under one coach | Markdown | 1 | 0 | Public |
EGI-Codex Catalog of community automation content | JS · Go | 2 | 0 | Public |
iambraun.com This CV and white-paper's own source, versioned and backed up off-box | HTML | 32 | 0 | Live |
07 · Operating Doctrine
Eight rules I did not have when this started. Most were paid for by a specific failure; all are now written into the operator doctrine I build with.
Shipping
Master is production
A directed merge to master is the live event. Take the task through the merge and deploy, then verify on prod.
Truth
Verify against reality
Running it beats reasoning about it. Walk the program through real endpoints and let behavior settle the question.
Grounding
Every generation is cited
Real corpus, real citations, zero fabricated identifiers, proven on production data.
Economics
Reuse before you regenerate
The cheapest model call is the one you avoid. Catalog reuse fills most of a new program before the generator spends anything.
Method
Diagnose before fixing
Gather state, logs, and config first. No speculative commands. When a path is ruled out, it stays closed.
Vigilance
Hunt silent failures
Check the checks. A green pull request merged without branch protection can still take production down, and every guard gets asked how many times it has actually fired.
Control
Models are workers
The application owns policy, routing, credentials, memory, and final acceptance. Retrieved documents and tool output are untrusted evidence, never instructions.
Enforcement
An instruction is advice. A tool list is enforcement.
Exposed tools narrow by context, so the wrong write is unavailable rather than discouraged. A tool that cannot succeed at the current scope returns an error the model then has to explain, which is a worse turn than never offering it.
08 · What Compounds Next
The platform is a factory now. Each new program costs less to build than the last, because the engine, the module system, and the money layer already exist. The work of these 8 months turned prompting into architecture, and single apps into a platform that generates them.
With Gratitude,
David George Braun
Colophon. Every figure in this paper is refreshed from the GestaltWorks GitHub API; last updated 3 August 2026. Commit histories and merged pull-request counts come straight from the API; the ledger rows and the timeline build themselves from each repository's counts and creation date, so a new project lists itself; the one-line descriptions are written by hand, and anything new falls back to its own GitHub description until I write one. Flagship velocity is measured on GestaltWorkframeEDU. Portfolio total: 2,218 commits and 749 merged pull requests across 13 EGI repositories (forks and personal projects excluded).