Not build-committed. This is the ACTUAL system the shop ran for years — the reference model the AI-orchestrated platform generalizes. Captured because it is the domain expertise.
The workflow (ticket-driven, osTicket)
Everything ran through one ticket. Canned per-department responses gated each transition: 1. Intake. Drop-off on the counter (later: locker drop). Remote CSR (Katrina, in the Philippines) via camera + remote-desktop to the CS machine: customer opens their account, CSR photographs everything together — contract + ID — and uploads to the ticket so the customer gets a copy too. (Transparency = trust; kept "weirdos" out.) 2. Estimate + consent. Ticket sends the estimate with a link; customer replies "yes I agree, repair my computer." Only THEN charge the card. 3. Parts. Order the part → "part ordered" status → when it arrives, "waiting for part" → check it in → notify "soon to be attended to." 4. Repair (Windows, the common path). Boot off USB → copy files off → erase → reinstall Windows (initial setup done FIRST) → blindly drag one folder named the ticket number back → boot to the login screen. Never see a filename. 5. Proof. Photograph the working login screen → post to the ticket (customer sees it's fixed). Wrap the machine in plastic so everything stays together. 6. Notifications. Running behind / started / done + the bill. 7. Pickup. Pay the bill → receive the locker combination to retrieve it. Payment-gated, no loitering in the shop.
Why it worked (the principles, all field-proven)
- Expertise in the SYSTEM, not the worker. Proceduralized enough that high-school
employees ("mostly hardware") and a remote CSR could execute reliably. THIS is the platform: software+process hold the judgment; the hands can be less-skilled or remote/vetted.
- Blind data handling. Ticket-number folder, no filename exposure → customer privacy AND
worker protection (zero exposure to disturbing/illegal content).
- Consent as a gate. "Yes I agree, repair my computer" before any charge/work.
- Provenance built in. Photos in the ticket (working screen, contract+ID) = the proof
trail, shared with the customer.
- Payment escrow. Locker combo released on payment.
- The ethos. "I've never been curious about people's sock drawers — don't look, you don't
have to remember." AND the skill to investigate (employee-usage-policy monitoring) used ONLY per policy/authorization. The power to look; the discipline not to. That's the whole consent/guardrail stance of this engagement, lived.
Maps 1:1 onto what we designed today
| Shop practice | Today's design | |---|---| | "Yes I agree, repair my computer" link | consent gate (fail-closed) | | Ticket trail + completion/ID photos | WORM audit-log + screenshot provenance | | Ticket-number blind folder drag | blind data-handling (handle-without-inspecting) | | osTicket status flow (estimate→approved→part→…→paid→pickup) | intent-first + status transitions + reconcile | | High-schoolers/remote CSR run the procedure | vetted-technician dispatch; expertise in the software | | Locker combo on payment | payment-gated completion / escrow | | Photograph the fixed login screen | two-oracle verify (visual proof of the done state) |
Implication
The hardware-inventory library crate (Sprint 2.9) is ONE component. The larger platform is this operational model, generalized: AI + software hold the expertise (remote, sight-independent), vetted local technicians are the hands, companies/customers move through the same consent-gated, privacy-preserving, provenance-logged ticket flow the shop already proved works.
Stack + build discipline ("everything thing," done right)
- The stack is TINY: Postgres + Rust/Forge + FreeSWITCH + the library crates already built.
Three pieces + an owned library. The breadth is in the CAPABILITIES, not the tech — that's good architecture, not sprawl (sprawl = many languages/services; this = one coherent foundation).
- Build INCREMENTALLY: spine (USD + ticket + DB) + ONE capability at a time, ship, add next.
Hardware-analysis (Sprint 2.9) = capability #1. The gates enforce this; it's what stops an "everything platform" from becoming vaporware. Never a boil-the-ocean rewrite.
- Telephony = self-hosted FreeSWITCH/FusionPBX for inbound calls: known, owned (no per-minute
rent — fits the no-Twilio-rent stance), capable. "Something better" candidate = SignalWire (FreeSWITCH creator's programmable cloud, same DNA) IF cloud/programmable voice is ever wanted. Solved, swappable component — don't over-optimize.
Technical infrastructure (platform components + a security capability)
The shop's technical spine, to reimplement/orchestrate on Forge (compose, don't rebuild):
- Imaging / provisioning: boot-stick images + PXE boot (network-boot a machine for
test/wipe/recovery without a USB) — the bootusb lineage.
- Drive station: test · data recovery (Linux/
ddrescue) · wipe (WipeStatus/
WipeMethod already in operations-drive-health).
- Per-device network isolation + behavioral listener (security capability, sharp idea):
every untrusted machine put on the network gets its OWN isolated segment (VLAN / netns / isolated bridge + own DHCP) so it can't touch other machines/customer data/shop infra (blast radius zero). A passive listener on that segment (Zeek/Suricata/structured tcpdump) watches behavior and flags "acting wrong": C2 beaconing, exfil, port-scanning, DGA/bad-DNS. In a repair shop the machines are often INFECTED (that's why they're in) — the traffic IS the diagnosis. Extends operations-net-discovery + infrastructure-security-scan.
- GUARDRAIL: THREAT monitoring on YOUR isolated segment, machine in your custody —
watch BEHAVIOR (is it phoning home to malware?), never read the customer's CONTENT. Detect the C2 callout; don't snoop the drive. Same blind-handling line. Disclosed.
- Selling point: "untrusted machines handled safely by default — isolated, monitored,
and if yours is infected we'll know AND have the evidence."
Remote management / delivery layer (RMM — field-proven, bash)
- Managed-endpoint provisioning customizes the install to open an OUTBOUND reverse-SSH
tunnel to a hardened phone-home server (pf-locked-down; only the operator gets in). Outbound-only = works behind ANY NAT/firewall with ZERO inbound holes on the customer side: the endpoint reaches out, you reach back through the tunnel. Whole attack surface collapses to one hardened box.
- Access over the tunnel: SSH terminal (primary — "all I need is a terminal"), remote
desktop (X / RustDesk).
- Business model: monthly managed-services subscription (MSP recurring revenue), not
one-off repairs.
- Delivery layer for today's design: the health monitor + stolen-device phone home via
this exact reverse-tunnel; remote diagnostics + actuation run over it. The shop-in-a-box ships through this, billed monthly.
- Modern equivalents (market validation the pattern is real): Tailscale / WireGuard,
Cloudflare Tunnel / ngrok (the reverse tunnel productized); TacticalRMM / Atera / NinjaOne / ConnectWise (phone-home + terminal + MSP subscription productized). The bash RMM was early + right.
- Terminal-first + AI = sight-independent: the operator's remote-support career was
terminal-centric; AI-in-a-terminal (this session's shape) reads output + runs commands, so managed services can run again remotely without squinting at small detail.
- GUARDRAIL: consented managed access (customer subscribed), on the operator's hardened
infra, WORM-logged. The phone-home server is keys-to-the-kingdom — its hardening + audit is the core responsibility.
USD — Universal Service Daemon (the core agent / client spine)
One minimal daemon on every managed machine — OUTBOUND-ONLY, NO open ports, phones home to a hardened control plane, orchestrated by "the robot" (automation/AI, not manual login) — that delivers UNIVERSAL SERVICE: health monitoring, diagnostics, remote support, actuation, and stolen-device phone-home all ride the same secure channel. One daemon, universal service, zero LAN attack surface.
- Why zero open ports matters: the agent is NOT itself an attack surface. Most RMM agents run
a listener that gets exploited (Kaseya-style supply-chain vector). Outbound-only + no listeners = the endpoint is invisible/unreachable on the LAN. Better than the billion-dollar products.
- Robot-orchestrated: desired-state control plane, not per-machine manual access — the seat
the AI now occupies (manage the fleet over the tunnels, tireless, scripted).
- Trust concentrates in ONE place: outbound-only ≠ powerless (the control plane can do
anything through the tunnel). So harden the control plane, gate on consent (subscription), WORM-log every action. Single responsibility = the elegance.
- Forge build: an agent library crate composing
infrastructure-jobs(worker/desired-state),
the health library crates, foundation-audit-log (WORM), + the reverse-tunnel transport. Every platform capability is a USD-delivered service, not a separate agent.
#### USD modern stack (the Forge/Rust rebuild of the bash RMM)
- Transport = WebSockets (evolves the reverse-SSH tunnel): outbound to :443, looks like
web traffic, tunnels through any proxy/NAT/captive firewall raw SSH can't — SAME zero-inbound surface (client opens out, listens on nothing). Still carries SSH / a terminal INSIDE it, so "all I need is a terminal" is preserved. NOTE: this is exactly the bridge used all session (Claude → server → WS → extension → browser); the USD pattern is already the runtime.
- Rust = one memory-safe static binary, cross-platform, no runtime — the secure agent bash
couldn't be. (The Forge framework.)
- WASM = sandboxed, portable capabilities. Each USD service runs as a WASM module the daemon
executes in a sandbox it can't escape (a buggy/hostile capability can't own the host); runs anywhere incl. a browser control-plane UI. Framework already uses WASM (the mouse humanizer).
- Graphing =
operations-metrics/foundation-telemetry(Counter/Gauge) → dashboards:
the health time-series rendered = the customer-facing value ("your fleet health over time").
- Monetization: the hardware-analysis system (Sprint 2.9) is ONE PAID CAPABILITY the USD
delivers — a service on the daemon, billed in the subscription, graphed on the dashboard. Full business assembled: outbound-only secure agent → managed subscription → capabilities (health/diagnostics/support/recovery) → dashboards.
- Auth = mutual TLS (mTLS) with per-device client certs (
wss://+ client cert): - Both sides prove identity: endpoint verifies the control plane (no MITM/impostor), control
plane verifies the endpoint's cert (only enrolled machines connect; NO shared secret to steal).
- The certificate IS the enrollment. Issue cert = enroll machine (cert +
MachineFingerprint
+ WORM AssetEnrolled, one act). Revoke cert = the kill switch — stolen/decommissioned machine instantly loses phone-home/management/access.
- Cert-authenticated channel ⇒ WORM entries are cryptographically attributable to a SPECIFIC
machine ("endpoint FZFT0Z3 ran this wipe" is proven, not claimed).
- Cost: a PKI (CA/issuance/rotation/revocation via CRL/OCSP) — real ops weight, solved pattern,
rustls in Rust. Ties to ADR-0017 fingerprint + Sprint 1.3 node-lock (per-machine crypto id already in-flight) — wire the cert to the fingerprint that exists.
AI-in-the-loop + promote-to-Rust ratchet (the operations intelligence layer)
The USD collects data → ships it via infrastructure-ai connectors (anthropic/openrouter/ollama + the ask-ai bridge) to the AI. Three tiers with a ratchet between them: 1. Deterministic Rust procedures — the known cases. Fast, TDD'd, cheap, no AI, no hallucination risk. The core. 2. Bounded AI autonomy — the "choices we're allowed to do": AI acts within the consent/budget policy (e.g. auto-order the fan under $X). WORM-logged. 3. AI on the novel long-tail — "things we haven't thought about": analyze, propose/guide, or escalate to the human when beyond bounds.
- The ratchet (the point): when the AI solves a novel case, that solution is PROMOTED into a
formal, tested Rust procedure → graduates to Tier 1. The frontier SHRINKS; the system gets more deterministic, more reliable, cheaper over time; the AI always works the shrinking unknown. AI = R&D; Rust = production.
- Why it's safe (guardrails already in place):
- HUMAN-IN-THE-LOOP AT PROMOTION — the architect reviews before an AI solution becomes law;
it graduates through the SAME gates (TDD, mutation-proof, 3-AI consensus). AI never silently bakes itself in. An AI fix must EARN determinism.
- Bounded autonomy + WORM — acts only within allowed choices; every decision logged, so you
promote from a real record, not a vibe.
- It's the harvest ratchet, applied to operations: harvest lifts a proven pattern from N
hand-rolled apps → one library crate; this lifts a proven solution from N AI-handled incidents → one Rust procedure. Same discipline, new domain. Nothing permanent is ever "the AI said so."
Self-service + AI-coached field service (the full loop)
Flow: customer push-button self-diagnose (from the server/portal) → USD collects data → infrastructure-ai connectors ship it to the AI for analysis (operator doesn't have to) → auto-order the part (bounded autonomy) → dispatch a local tech → AI coaches the tech through the repair in real time (reads him the facts + diagnostic + prior conversations + proven procedures until he knows what he's doing).
- Why it scales (the moat): AI coaching DROPS THE SKILL BAR for the physical labor. Without
it you need expert techs in every city (rare/expensive/unstaffable). With it you need mechanically-capable hands the AI can guide — orders of magnitude larger labor pool, lower cost. Expertise lives in AI + procedures, DELIVERED on-site in real time. Every capable person in every city becomes a potential technician.
- Coaching from the record, not vibes: the AI reads the WORM ticket history + this machine's
diagnostic data + prior conversations + proven procedures. Facts, cited.
- Frees the operator: no analysis, no hand-holding — the system + AI run ops. The architect
designed it; isn't the bottleneck. Sight-independent, scales past one person.
- Guardrail: AI coaches WITHIN known/proven procedures; tech confirms each step with photos
(two-oracle proof, = the "photograph the fixed login screen" practice); novel/risky → escalate to a human. AI advises; the human tech executes and can refuse. Bounded like all autonomy.
- The point: it doesn't replace the operator's expertise — it MULTIPLIES it, present in
every city at once, for customers/techs they'll never meet.
Telephony / CSR layer (FusionPBX) — humans augmented, not replaced
- FusionPBX (FreeSWITCH, open-source, operator knows it) = the phone system. Compose, don't
rebuild (like application-tickets for ticketing).
- Call ↔ ticket, both ways: inbound call screen-pops the customer's ticket + machine health
(the CSR knows who + what's wrong before "hello"); call creates/attaches a ticket; click-to-call from a ticket; voicemail → transcript in the ticket.
- AI on the phones: AI voice agent for after-hours/overflow/IVR triage; live transcription +
summarization back to the ticket; real-time context fed to the human CSR mid-call.
- The humane pattern (core to this platform): it EMPLOYS people — remote CSR (Katrina) on the
phones, capable local techs on the hands — real jobs, coordinated + AUGMENTED by AI, not eliminated. AI gives the CSR instant context and the tech an expert-in-his-ear; it fires neither. AI multiplies humans and CREATES work; one architect holds the whole thing.
Business model (the real product): vertical SaaS — "shop-in-a-box" for repair
- Customer = the capable-hands, software-underserved tech/shop. Great mechanically, no time
or wiring to build ticketing/ledger/parts/privacy automation. They never had a Dan. That gap is the market. The "vetted technicians" are CUSTOMERS/partners, not just dispatched labor.
- Founder fit (rare, non-fakeable): the trade knowledge AND the proven business system,
both seats lived. A coder without a shop can't replicate it; a bench tech never built the ledger.
- Category: vertical SaaS for the trades — ServiceTitan (HVAC) / Toast (restaurants) /
Jobber (home services), but for independent COMPUTER repair, which nobody has really done — with privacy-first / blind-handling / consent-gated DNA built in.
- Why buildable now by a tiny team: AI writes the software and does the looking, so the
founder occupies the sight-independent systems+vision seat. The thing proven today.
- Moat: NOT the code (code is cheap now) — the lived operating system + customer/tech trust
+ onboarding. "Here's how a real shop protects the customer AND the tech, proven over years."
- What it packages: the osTicket-style flow (intake→estimate→consent→pay→parts→status→
proof photo→locker/escrow) + hardware-inventory/health/diagnostics + WORM provenance + the consent/transparency guardrails. All the research notes in this dir are its component specs.