How we work
How we build software, and real work that shows it.
What we believe, in practice
Six principles, each one enforced by how the software is built rather than by a policy document
Keyless where we can be
Where Microsoft and GitHub allow it, a workload proves who it is with a short-lived token instead of a stored password or key: our sites send mail this way, and every CI run gets its own token. Some keys remain. The ones applications and CI read are named on a reviewed list that holds names only, never values.
The session lives in your browser
Sign-in is Microsoft Entra ID, handled in the browser. By default the token lives in that tab and is gone when the tab closes, unless you choose to stay signed in across tabs. Admin-level tokens are held in memory for the one action that needs them, then thrown away. There is no server-side session for us to store, leak, or forget to expire.
Your data stays in your tenant
Our compliance product keeps every answer, upload, and attestation in a SharePoint site inside the customer’s own Microsoft 365 tenant, reached with the user’s own permissions. Your tenant data never reaches us, so we cannot misuse it. What we hold about you is billing, nothing beyond what Stripe, our payment processor, holds.
Built once, used everywhere
Identity, consent, security headers, forms, mail, theming, translation, testing, and lint rules are among 46 versioned packages every application consumes. A fix ships once and reaches every app, with CI proving each landing.
Green means go
Nothing merges until the automated checks pass in the merge queue, and no one merges past a failing check. When the fast checks are green, the change merges into the preview branch. A person decides when a release candidate ships to production.
We run on what we build
This site, our other sites, and our tools all run on the same packages and the same pipeline. We use what we build every day, so we find the rough edges first and fix them upstream, once, for everyone.
One change, end to end
A real change, from ticket to production, with the timestamps
Every change we make goes through the same gate, whether it is one line or forty thousand. Here is a real one from this website, with the timestamps GitHub recorded. It is from before the preview lane, when a merge went straight to production; the path today is told once, under Engineering below, and its latest release is shown after this one.
- The ticket
- Issue #446 asked to remove the Fleet Dashboard from this site, since the dashboard had moved to its own application. It states what goes, what stays and why, and what is out of scope, before any code changes.
- The change
- Pull request #447: 233 files, 198 lines added and 41,894 removed. The dashboard had moved to its own application, so this deleted the old copy, including every line of sign-in code. The public site now has nothing behind a login.
- The gate
- All seven required checks passed in the merge queue against the exact commit that would land, 23:20:18 to 23:22:52 UTC.
- The merge
- Squash-merged by the pipeline at 23:23:12 UTC on 2026-09-09, 2 hours 28 minutes after the pull request opened. The linked issue closed itself two seconds later.
- The deploy
- The post-merge run built and deployed production from 23:23:16 to 23:25:25 UTC. Merge to live took about two minutes.
Nothing is hidden in the trail. The ticket and the pull request were opened together, as one piece of work. The whole trail, from the reasoning to the checks to the deploy log, is in the repository for anyone we give access to.
- The release
- Pull request #893, “Release preview to production”, from the promotion branch promote/next into main: opened 15:59:05 UTC and merged 16:33:19 UTC, 34 minutes later. It carried what had merged into preview since the last release.
- Where to read it
- #893 (opens in a new tab)
- Pull request
- #447 (opens in a new tab)
Where everything lives
The architecture, as who holds what; the picture is on the home page
Most of what makes an application risky is what it keeps: sessions, keys, customer records. We design ours to keep as little as possible, and put each thing that has to exist with the party that already owns it.
- The map, and what each side holds
- Where your data lives
One requirement, every framework
A real control from our compliance registry, mapped across frameworks
Our compliance registry links statutes and standards into one graph, so a single real-world control is recognized under every framework that asks for it. Multi-factor authentication is a good example. One question in the registry is linked to all of these:
| Framework | Identifier | Requirement, as the registry names it | Linked |
|---|---|---|---|
| NIST SP 800-171 Rev 3 | 3.5 | Identification and Authentication Requirements | Directly |
| NIST SP 800-53 Rev 5 | IA-2 | Authentication | Directly |
| ISO/IEC 27001:2022 | A.8.5 | Secure authentication | Directly |
| NIST CSF 2.0 | PR.AC | Identity Management, Authentication, and Access Control | Directly |
| HIPAA Security Rule | 164.312(d) | Person or entity authentication | Through ISO A.8.5 |
| NY DFS 23 NYCRR 500 | 500.12 | Multi-Factor Authentication and Access Controls | Through NIST 800-171 and 800-53 |
The question itself. “An admin argues that long-tenured staff and service accounts should be exempt from multi-factor authentication (MFA) because they’re ‘trusted’ and it would slow them down. What’s the strongest response?” The correct answer explains that exempting any account creates a single point of failure. The three wrong answers are the arguments people actually make.
Why it matters. Answer once, and the evidence counts toward each of those requirements. That is the point of a crosswalk. This question is still marked draft in the registry while the content owners review it, and the registry holds NIST 800-171 at family level (3.5), not as individual requirements.
The rules we write for ourselves
The rules we hold ourselves to, and a real decision record
Every repository carries the rules its contributors follow, human or AI agent, as files loaded at the start of every session. The first three rules every agent reads are:
- Never fabricate. No invented numbers, results, or checks reported as passing that were not run.
- Show your work. Every claim states where it came from: the file and line, the command, or the check.
- Mark what’s unverified. Anything that could not be checked says so, in the output itself, every time.
Lean is a decision, and we write it down. We also cut what does not pay for itself. A CI check that confirmed each pull request linked an issue did five seconds of work and billed a full minute per run. We deleted it (Platform-Packages #481, filed in that repository before it moved to Turbo-Shared-Services/Packages, where the number does not resolve), wrote down what enforcement it gave up, and kept the rule as a convention. Every decision like that is recorded the same way: what we did, why, and what it cost.
Engineering
The platform, the pipeline, security and how we run a project, each told once
Platform architecture
One platform repository publishes 46 versioned packages and the shared CI workflow, and every site and tool consumes both. The packages are ES modules with type declarations, published privately to GitHub Packages with Changesets, and consumed as ordinary dependencies, so a fix ships once and every application picks it up through Dependabot.
Identity. Sign-in runs in the browser with Microsoft’s MSAL library. By default the token is kept in that tab’s session storage; a person can choose to stay signed in across tabs. A rare admin-only action, such as provisioning a SharePoint site, uses a second sign-in held only in memory, with a warning banner on screen while it is live, and discards it afterward. Microsoft endpoints for the commercial, GCC High, and DoD clouds are a typed table in one package, and an unknown cloud fails the build.
Configuration storage in the customer’s tenant. The Portal provisions one SharePoint site per customer tenant, the workspace map, holding these lists: Tenant, Products, ProductSettings, Groups, GroupRoles, GroupRoleHistory, ProductInstalls, ManagedTenants, and UserPreferences. Each product serves a JSON manifest at a well-known path declaring the lists, fields, roles, and settings it needs, and reaches its own site through the Sites.Selected Graph permission. A product decides what a person may do from their Entra group membership.
Application shell. Each Next.js application applies a shared configuration that sets Content Security Policy, HSTS with preload, X-Frame-Options, nosniff, a strict referrer policy, and a Permissions-Policy, and mounts Vercel BotID for public endpoints. The consent package gates storage, statistics, and marketing behind a choice the visitor can refuse, a lint rule forbids vendor endpoints written outside the package that owns the vendor, and an opt-in rule forbids tracking that bypasses consent. The UI package is themed through 20 semantic colour roles with light, dark, and high-contrast settings. Our sites send mail through Microsoft Graph with no stored credential, and the payments package refuses Stripe live mode unless it is switched on explicitly.
Hosting. Vercel, one project per repository, deployed only from GitHub Actions; Vercel’s own Git integration is off.
Engineering methodology
Every site and tool repository ships through one reusable GitHub Actions workflow, one required-check set, and one merge queue. The workflow is defined once in the platform repository and reaches the other repositories through a tag that moves only after a canary repository passes on the new version. The rules every contributor follows, human or agent, are files in each repository, loaded at the start of every session.
The delivery pipeline. Main requires seven checks to succeed before anything lands there: Generate Artifacts, Security Scan (Gitleaks credential scanning plus actionlint on the workflows), Dependency Rules (import-graph boundaries and cycles), Lint & Format (oxlint and Prettier), Type Check (TypeScript 7, with `any` banned), Unit Testing (Vitest), and Build (a real Next.js build). A pull request into preview needs the first four of them (Security Scan, Dependency Rules, Lint & Format, Type Check), and the merge queue runs them against the exact code that would land. When they are green, the machine merges the change into preview.
- Labeled issue
- Branch and pull request
- Merge queue runs the fast checks
- Squash into preview (this site has no preview host: the fast checks run and nothing deploys)
- Daily release candidate runs the full suite
- A person merges the release PR into main; main deploys to production
This site has no preview host, so a merge into preview runs the fast checks and deploys nothing. At 10:00 New York time a release candidate is cut from preview and runs the full suite. Production changes only when a person merges the standing release pull request into main, as a merge commit; in the platform repository a person also merges each versioned package release. Main accepts pull requests only from preview, promotion branches and Dependabot, and it deploys from GitHub Actions in the same run, with the same install step, as the checks.
How we decide what to build and how
- One stack. The standard is TypeScript 7, Next.js, React, Tailwind, oxlint, Prettier, Vitest, and pnpm.
- Duplication over the wrong abstraction. Code moves into a shared package when it reaches its second or third real use, and the platform repository keeps a ledger of what has moved.
- Change forward, never roll back. A bad deploy gets a forward fix through the same checks, including a revert of the merge in a new pull request. No one merges past a failing, pending, or skipped check.
- Every change links a labeled issue. Issues carry the same label keys in every repository (size, type, needs, priority, status, env). The needs label says what an issue waits on: an agent, a decision, a look, hands at a keyboard, or money. There is no organization-wide board; the issues are the record.
Dependency hygiene. Dependabot runs in every site and tool repository. Consumer repositories merge green updates automatically; the platform repository batches them on one branch and merges them on purpose, because a bad update there reaches every application.
Security, identity, and secrets posture
Workload identity. Our sites send mail with no stored credential: each production deployment exchanges a Vercel OIDC token for a Microsoft Graph token as its own Entra app registration, and Mail.Send is granted only through Exchange scoping to that site’s own mailbox. GitHub Actions uses a token minted for each run.
Keys that remain. Some keys still exist. The reviewed key list places eight in GitHub’s organization secret store for CI: two for the deploy and the CI app, and six test-account credentials shared only with the repositories that run browser tests. Applications read four at runtime from sensitive Vercel variables. The list holds names only, never values, and the stores themselves were not re-read for this page. Keys are rotated by hand on a calendar until rotation is automated.
Customer data. Governance keeps answers and evidence in a SharePoint site inside the customer’s tenant, reached with the signed-in person’s own delegated token. Lint rules forbid server actions and any route that reads a request body.
Pipeline and application controls. Gitleaks scans every change for credentials and actionlint checks the workflows themselves. `any` is banned. Every web application ships the same security headers from one shared configuration, and public endpoints carry bot verification.
Our approach to a project
- 1. Discovery and planning. You talk to the people writing the code. There is no account layer between you and the engineer who understands your constraints, and no rewriting of what you said on the way through.
- 2. Design and architecture. We design so the guarantee holds structurally, not procedurally. Governance keeps evidence in the customer’s own Microsoft 365 tenant, because a promise about your data is only worth what the architecture enforces. What we hold is billing, nothing beyond what Stripe, our payment processor, holds, plus a short list of named exceptions, which the privacy notice lists.
What it runs on
We run no database of customer data. These are the services the Tools and this site depend on, what each one does, and how far it reaches.
| Service | What it does for us, and how far it reaches | Why we use it |
|---|---|---|
| Vercel (hosting and delivery) | Serves the code of this site and the Tools from its network, and runs a short list of server requests: the tools menu, bug reports, health checks, the Portal’s billing check, this site’s contact form and Bid’s scheduled jobs over public federal data. It stores Bid’s public feed, checks forms for bots, and measures visits to this site after you consent. It logs each request’s network address and the page asked for; how long it keeps those logs is not yet confirmed. Your tenant’s data does not pass through it, apart from what you type into a bug report. | One platform builds, deploys and serves every site and Tool. Production changes when a person merges the release pull request, as the pipeline above says. |
| Stripe (billing) | Takes payment and keeps the billing record, including which Tools you buy, as our payment processor. Our code only reads a checkout and its subscription; it never writes to Stripe. | Billing is the one record the sale needs, and Stripe keeps it so that we do not. |
| GitHub (code, checks and bug reports) | Holds our source code, runs every check before a change merges, publishes our shared packages, and receives the bug reports you choose to send, as issues in an internal repository. | Every change, check and report is recorded in one place. |
| Microsoft 365, our own tenant (mail) | Sends this site’s contact-form mail and holds inquiries in our shared inbox. The site gets a short-lived token for one mailbox and keeps no password. | We run on the same Microsoft platform we build for. |
| SAM.gov (public contract data) | Bid reads public federal notices from it. Nothing about you is sent to it. | It is the government’s own source for federal notices. |
| GoDaddy (DNS) | Answers which server each of our addresses points to. It sees no Tool traffic. | Our domain’s name servers are GoDaddy’s. |
What it is built with
The main frameworks, not every package.
- Next.js and React. Every site and web Tool is a Next.js application, written in TypeScript with React. Most of each page runs in your browser.
- Microsoft Authentication Library (MSAL.js). Microsoft’s own sign-in library for browsers handles every sign-in and keeps the token in your browser.
- Tailwind CSS. Styles every page. It runs when we build, not in your browser.
- Tauri and Rust. The Windows desktop app, Telemetry, is a Tauri application: Rust reads Windows, and a React screen shows it.
Upcoming: GCC High and DoD tenants
The apps already know the U.S. government cloud’s sign-in and Microsoft Graph addresses. The app registrations in that cloud do not exist yet, so a GCC High or DoD tenant cannot use the Tools today.
When they do, we plan to serve those editions from Azure Government instead of Vercel. Azure Government holds a FedRAMP High authorization; Vercel is not listed in FedRAMP’s published Marketplace data (last updated May 19, 2026).
DFARS 252.204-7012 requires a cloud provider that stores, processes or transmits covered defense information for a contractor to meet the FedRAMP Moderate baseline. The Tools keep that information in your tenant, and moving their delivery into Azure Government puts the code they run in an authorized cloud too. The Tools support your CMMC work; they do not certify you.
Check us yourself
You do not have to take our word for any of it. A technology-detection browser extension, such as Wappalyzer, lists the frameworks, hosts and analytics a website loads. Run one on this site, on our Tools, and on every other vendor you are weighing. If you would rather not install an extension, your browser’s developer tools show the same thing on the Network tab.
What a browser recorded on October 26, 2026: before you accept statistics, this site loads nothing from any address but its own. After you accept, it adds one script, Vercel Web Analytics, also from its own address. Our Tools load no analytics at all.
Signing in to a Tool takes you to Microsoft’s sign-in page. That page is Microsoft’s, not ours, and it sends Microsoft’s own telemetry to browser.events.data.microsoft.com. We do not control it, and you will see it on any Tool that signs in with Microsoft.
On any vendor’s site, count the analytics and tracking tools a page loads. Each one is another company that sees your visit.
Prime or partner
How we fit on a contract
Who we are
We are a Service-Disabled Veteran-Owned Small Business (SDVOSB) and a certified HUBZone small business. Team members hold active security clearances (confirmed August 4, 2026).
What you get
Repositories you can read end to end, the issue trail behind every change, and TypeScript on Next.js with Microsoft Entra sign-in, which your own engineers can maintain.
Want to see it for yourself?
Ask by message and we will reply by email about walking you through any of these samples in the actual repositories.
