Five acts, operation to boardroom

The brief said "a process you know well." I built theirs.

Three reasons this is my case study — the exact three the brief asked for.

Worked with it

I built their ticketing system

I've been building Flash Dental's service-ticketing — which put me inside their after-sales operation, on the very documents you'll see next.

Ripe for it

Real mess, real money

Capital equipment, an archipelago, run on hand-typed sheets and WhatsApp. The friction is real and the leak is measurable — I've watched it happen.

Newly possible

What I'd explore next

The mess is exactly what modern document-AI is finally good at. This isn't the ticketing tool I built — it's how I'd modernize the whole process with AI.

Who's on the other end — a clinician whose room goes dark when a unit is down. The operation runs for her uptime.

01

The operation I know

Flash Dental sells and rents dental equipment across Indonesia, then services it for years. I've been building their ticketing system — so I know this operation from the inside; this case study runs on their real data.

Capital gear — sold or rentedDental units, sterilizers, compressors, imaging (Esay, DTE, Kuraray) — bought or rented from Flash's own fleet, its gear to track & service.

Nationwide, few handsJakarta to Ketapang, Tarakan, Medan — a lean field team (Yudi, Dedi, Febri, Arga…) covering the whole archipelago.

A lifecycle, not just repairsInstall → scheduled maintenance → breakdown service → survey → training; on-site and remote.

After-sales is the relationshipA unit down = a treatment room offline. Parts, speed, reliability are the recurring, high-margin business.

A Flash technician on-site — the after-sales work that keeps a chair running

A clinic reports a fault; the fix is rarely same-day. The unit waits — often days to weeks — while a spare part is checked, ordered, and shipped across the archipelago.

Their real service log — a sample of the actual rows (translated from Bahasa). Machines, jobs and technicians all tracked by hand in Google Sheets, and the mess is right in the cells.

Modeled conservatively from their logs, in USD for scale (≈ Rp 16,000 / $1). This is Flash's loss — not the clinic's downtime — and their brief admits none of it is measured today.

Not incompetence — the whole operation is one person babysitting a spreadsheet, and no tool could ever sit on top of the mess.

An operation you can't measure quietly devalues the business.

Because it's manual and unmeasured, Flash silently breaks its own promises — the six-month maintenance, the honored warranty, the fast response — and can't see its service quality, its people, or its product reliability. That's the ~$170k a year from the last slide, plus the enterprise value a measured operation is worth. Incumbent BI & EPM assume clean data these operations don't have.

One engine, three layers — from the operation up to the boardroom.

Modernizing isn't "AI does everything"

The value is drawing the line in the right place — and keeping people where judgment, hands, and trust belong.

Scattered documents flow left to right — parsed, indexed, reasoned over, served, measured, and governed. Every slide that follows zooms into one of these six.

One pass over the raw documents produces both the searchable index and the KPI tables.

Extract → de-identify → chunk → embed. The unstructured half feeds retrieval; the structured half (dates, amounts, reps, statuses) feeds the computed KPIs — no clean warehouse required upstream.

One scanned form → structured, computed, masked

A real row — Bahasa, handwritten, stamped — becomes typed, structured, warranty-computed, PII-masked signal in one pass. Low confidence would set _needs_review: true and route it to a human — never silently ingested.

Google Cloud–native — the same six layers from the backbone, chosen for the mess, the residency, and the cost curve.

Assets and people aren't two solutions — they're two data domains in one UI, for the same C-suite. RBAC decides who sees what.

One service record → two measured truths

one service-log row"Yudi closed FD-1042 · suction motor · Surabaya · reported 19 Jun, done 29 Jun" — plus the linked sales invoice

→

Asset truth3rd suction failure → reliability; invoice date → warranty status & age → maintenance due

Technician truthYudi closed it → jobs done, first-time-fix, cycle time, out-of-town load

The manager's dashboard and the technician's job — the same platform, scoped by role, on the phone they already carry.

9:41 5G

Cockpit C-level

First-time-fix

80%

▲ 6 pts

Cycle time

7.8d

▼ 1.2d

Jobs · Jun

42

▲ 5

Outstanding

7

▼ 3 aging

SLA

94%

▲ 2

Warr. due 30d

12

FD-1042

First-time-fix by region

JKT

SBY

KLM

BDG

DPS

MDN

What needs you today

DC-300 fails near warranty end — 3 units, Kalimantan

5 jobs stuck on stockout — Selang & Footcover

Yudi carries 38% of out-of-town jobs

Job queue

FD-1042 · SurabayaOutstanding

drg. Iffah · JakselWaiting part

drg. Andi · NunukanPM overdue

drg. Lidya · BanjarbaruFinish

9:41 5G

Field Teknisi · Yudi

FD-1042 · DC-300

Suction failure — 3rd

Klinik Sehat · Surabaya · 2 days ago

⚠ Warranty expired Jun 2024 — billable

Guided diagnosis

Recurring cause is a clogged suction filter — replace it with the motor, or it returns.

cited: manual DC-300 · §suction

BringSuction motor + filter kit · check warehouse stock before travel.

Service history
2025-01 suction motor · replaced
2024-09 hydraulic leak · sealed
2024-03 suction motor · replaced

Next PMDue Aug 2025 · combine with this visit.

One platform, scoped by role — the C-level's dashboard and the technician's job are the same system, different view.

The manager, in the pocket — live first-time-fix, cycle time, the outstanding backlog, and the day's alerts.

The technician, at the chair — the unit's history and warranty before the toolbox opens; no 45-minute invoice hunt.

Grounded, cited answers on both — the copilot proposes; the person decides and acts.

The same engine behind the cockpit. Pick a domain, ask — it runs entirely in this page.

How it fails — and the guardrail

A system the C-suite steers by has to fail safely and visibly, never silently. Every failure mode has a designed defense.

Land one instance, prove it, generalize — never boil the ocean.

One engineer, AI-assisted. The build is days — the schedule is just honest about what actually gates it.

What v1 changes in the first two weeks

AI-built by one engineer, run lean — so the whole spend is a rounding error against the ~$170k a year the mess costs.

The hard questions — answered head-on

The tech finally reads the mess — and I already know the operation.

Why now — Gemini can finally read the mess (unstructured, Bahasa, cross-department documents) and manufacture the clean signal BI always assumed was already there; GCP's Jakarta region makes in-country data residency real. Neither was true two years ago. Why me — I've been building Flash Dental's ticketing system, so I know this operation from the inside, on its real documents — and a career of field-worker mobile tools (Invoice2go, Pawoon) and messy-data pipelines is this exact problem, over and over. Not a concept deck — the operation I know best, modernized.

Let's prove it on one operation — in a week or two.

v1 is a grounded copilot over one operation's real, messy documents. The mess becomes answerable, and we measure the lift — first-time-fix, cycle time, warranty leakage caught — before touching anything else. Three things to start: