← Back to work

Turning scattered company data into a decision hub — the warehouse, the backend, the AI layer and the interface, designed, built and shipped end to end by one person.

KPLUSOperations Hub

Data × AI × Full-stack
KPLUS Product design + full-stack build Internal decision hub · AI × data 2026

(ROLE)

Product design
+ full-stack build (solo)

(STACK)

BigQuery · Flask ·
Cloud Run · IAP

(SCOPE)

Data warehouse · backend ·
dashboard · design system

(TYPE)

Internal decision hub
(AI × data)

Most dashboards show everything and decide nothing. KPLUS needed the opposite — one place where the numbers that actually change a decision rise to the top. I designed and built the whole thing: the data warehouse, the backend, the AI layer and the interface.

Live · IAP-secured
Operations Hub — the system

One system, from scattered departmental exports to a single decision.

60tables in a layered warehouse — six departmental sources, designed and built by one person.
dim · factmartservingviewsdashboard
10views
6source systems
5decision views
1person · AI-assisted
The system at a glance — structural scale, not business figures; shown in the product's own brand blueScale
The visual system behind these screensdata-viz tokens
Neutral · basepaper / grey / ink
Primary · active series#4D7CFF
Light · secondary#86AAFF
Deep · emphasis#2A3FCA
Type — Inter for text · mono, tabular figuresCharts — one blue palette, ranked by intensity, no red / greenRestraint — neutral by default; blue is the only accent
Operations Hub — daily overviewNews tab · dark monitoring mode
ILLUSTRATIVE · dummy data
AI DAILY SCAN 3 signals today: two channels are behind on restock, one SKU has high try-on but low conversion. Suggest prioritising the store stockout first.
Sessions today
1,240+6%
Open alerts
3−2
Sources synced
6✓
Data freshness
2h
MonTueWedThuFriSatSun
ALERTS · ACTION NEEDED TODAY 3
ItemSignalStatus
Store B · stockoutrestock behindact
SKU-07 · high try-on, low conv.funnel gapwatch
Social · race recapreach above plangood
News tab — the AI daily scan reads the data, writes the takeaway and flags what matters (illustrative)The product itself
◆One person. The whole stack.

Two halves that rarely live in one person — the design and the engineering. On this project both are mine, end to end. Left is the interface and the design system; right is the real code behind the data, backend and deployment.

[ DESIGN ]interface · system
1,240Sessions
3Alerts
2hFresh
AaInter · Mono
= same person
[ ENGINEERING ]data · backend · deploy
# warehouse — one trusted number
CREATE VIEW v_sales_kpi AS
SELECT channel, SUM(net) AS rev
FROM serving GROUP BY channel;
# backend — live SQL, secured
@app.route("/q")
def q(): return run_bq(req.json["sql"])
# pipeline — repeatable & safe
def build_serving():
    backup()
    for s in SOURCES: upsert(load(s))
Cloud Run · IAP · domain-locked
Left: the interface and design system. Right: the real code behind data, backend and deploy. One person, end to end.One person · full stack
01The problem

The company's data lived in scattered CSVs, an ERP, POS machines and a handful of tools. Reports were complete but not decidable — leadership had plenty of numbers, but not the next action.

  • The daily report was re-uploaded by hand, every single day
  • Data spread across sources, with no single source of truth
  • Reports answered "what happened", never "so what do we do"
  • Too much on screen — the signal was buried in the noise

The deeper problem was coordination: retail, design, sales and product each held a piece — what was launching when, where it stood, what to push next — but no one saw the whole. My first version made it worse: I patched whichever team shouted loudest, one page at a time, and ended up chasing symptoms. So I stopped, laid every team's problems side by side, and rebuilt around one logic instead of another patch.

The brief I set myself: build a decision hub, not another report.

02The core principle

The whole product runs on one rule — optimise for decidability, not completeness. It's what every design decision is measured against.

Every block earns its place

For each number I ask: "whose action does this change?" If there's no answer, it gets collapsed or pushed to a deeper layer — never the main stage.

Titles state the conclusion

Not "weekend sales" but a headline that names the takeaway; not "try-on count" but one that surfaces the tension underneath. The title does the thinking, so the reader doesn't have to.

Effort → effect

Inputs (launches, campaigns, content) are aligned against outputs on one timeline, each point traceable to the real deliverable. The panel doesn't just show — it evaluates: was this worth doing again?

Any number
→
Whose decision does it change?
→
Yes → surfaced to the main stage
No → collapsed, pushed to a deeper layer
The filter every number passes — if it can't change a decision, it doesn't earn the stageDecision logic
GLANCEThe conclusion, in one line — the answer first.
EXPANDThe breakdown, opened only on demand.
SOURCEThe raw rows underneath — traceable, auditable.
Three layers of depth — conclusion, then detail, then the raw sourceInformation depth
03Architecture — built end to end

A spreadsheet could hold the data, but not carry what this needed — backtesting, cross-referencing, one auto-updated source everyone trusts. So I didn't stop at a sheet: I built the full stack, from the data model to the pixel.

01· SOURCES
ERP snapshotSales CSVInventoryShipmentsStore POSSocial APIs
Raw, scattered, inconsistent
02· PIPELINE
Python gen_*.pybuild_serving (row UPDATE)auto-backup before rundata validation
Repeatable & safe
03· WAREHOUSE
BigQuery serving tableSQL views (v_sales_kpi…)single source of truth
One trusted number
04· BACKEND
Flask on Cloud Run/data/q live SQLIAP · domain-locked
Secure by default
05· FRONTEND
Single-file UIauto-bind datainline fallbackdark / meeting mode
Never breaks on load
End-to-end: data model → pipeline → warehouse → API → interfaceBuilt & deployed solo
Security & ops by design: deployed on Cloud Run behind IAP and locked to the company domain — no anonymous access, re-verified on every deploy. Running a real internal tool means owning reliability and security, not just the visuals.
04Making the data decidable

Raw company data is messy, overlapping and sometimes missing. The hard, valuable part isn't drawing charts — it's the judgment that makes a number trustworthy enough to act on.

Revenue by channel — auditableevery figure traceable to source · illustrative
ILLUSTRATIVE · dummy data
DealerRetailKAExportOther {待}Direct
Grey bar = source pending, flagged, excluded from totals ≈90% cost coverage
The judgment made visible — primary = trusted, grey = pending & excluded, margin shown as an estimate (illustrative)Decidable data
Channel classification by code

Customers are grouped by customer code, not name — so messy, inconsistent names never corrupt the channel split.

Subset de-duplication

Overlapping channel groups are de-duplicated so revenue is never double-counted, and an exclusion list removes non-dealers.

Honest margins

Where exact cost is missing, margin is an estimate, clearly labelled (≈90% cost coverage) — never a fake precise number.

Single source of truth

Headline revenue is migrating from an ERP snapshot to one BigQuery view — so everyone argues from the same number.

No fabrication

If a number's source is unknown it's marked pending ({待}), greyed and collapsed — never filled with a guess.

Traceable to source

Every figure can be traced back to its CSV row or BigQuery view — the dashboard is auditable, not a black box.

◆Judgment in one view

Once the data is trustworthy, the real value is using it to reach a non-obvious call. On channels, I didn't rank "which one earns the most revenue." I asked the question that matters for a physical-goods brand —

Are the channels that move volume the same ones that make margin?

Margin × volume — every channel placedindexed · anonymised · position = profitability × units, solid = growing
ILLUSTRATIVE · indexed
Earns & moves · sweet spot
Volume, no margin · misallocated
Earns, small · scale up
Marginal
Own Aidx 58 ↑
Own Bidx 37 ↑
Overseasidx 62 ↑
Dealeridx 100 ↓
E-comidx 23 ↓
Staffidx 3
shift resources up-right
Margin (indexed) →
Volume · revenue index →
solid = growing dashed = shrinking circle size = revenue index · one blue palette, no red/green
Same data, a different question yields a different decision: the biggest channel (Dealer) is shrinking, the fastest-growing one sits in "volume, no margin", and the real profitable growth is owned retail. Conclusion — move resources up-right.Judgment, not charts
Decision · a chart I refused to ship
Situation
Building an "accounts receivable" view, I found every customer's AR exactly equalled their revenue — it was order value, not a true unpaid balance.
Options
A: ship a slick "AR radar" anyway — no one would notice. B: flag it pending and hide it from the decision surface.
The bet
I chose B, giving up a chart that would have looked great. A chart that misleads a decision is more dangerous than one fewer chart.
Second-order
This became the product's stance: every number is traceable and auditable. Leadership trusts the tool because it says "I don't know" when it doesn't.
05Solving the real problems with AI

The data was complete — but complete isn't useful. Leadership had no time to read tables; non-technical staff couldn't query them. AI is the layer that turns data into something a busy person acts on. That's the capability that matters most here — not "adding AI", but using AI to solve specific problems.

Raw, scattered data
→
AI layersummarise · answer in plain language · flag what matters
→
A busy human, deciding faster
AI isn't decoration — it's the layer that turns data into actionAI in the loop
PROBLEM
Leadership drowns in numbers — plenty of data, no "so what".
SOLVED WITH AI
An AI daily summary reads today's data first and writes the takeaway — the first thing leadership sees each morning.
PROBLEM
Non-technical staff can't write SQL to ask the data a question.
SOLVED WITH AI
An AI knowledge advisor — ask in plain language; it queries and answers.
Varied views — one blue system, different chart typesILLUSTRATIVE · dummy
Trend · line / area
Composition · donut
Dealer 45% Retail 30% Other 25%
Ranking · horizontal bars
SKU-07
SKU-02
SKU-15
SKU-21
◆Judgment isn't the happy path — it's deciding which failure matters

AI draws a perfect happy path in minutes. What's scarce isn't the screen — it's the judgment about where a number quietly breaks. So before designing each decision view, I hand the business rules to AI and ask it to break them — surface every edge case that distorts a figure. AI generates that failure map in minutes; my value is deciding which ones to guard, and how.

surfaced the edge case
My call → the guard
One customer code maps to two channels
Classify by code priority, never by name
Overlapping channel groups double-count revenue
Subset de-dup — revenue counted once
A SKU is missing its cost
Margin shown as a flagged estimate — never 0, never blank
Revenue from an unknown source
Marked {待}, greyed, excluded from totals
ERP snapshot and live BigQuery disagree
One source of truth (BQ view); migration noted
A non-dealer slips into the dealer group
An exclusion list removes it
Returns / negative-value rows
Use net, not gross — defined explicitly
Duplicate orders across POS and ERP
De-dup on the order key
A new SKU has no history
No fake trend — show "not enough data"
Data hasn't refreshed in hours
Show a freshness flag — never pass stale as live
AI widens the failure surface; the judgment — which failures matter and how to guard them — is mineThe judgment layer
Execution is going free: drawing ten versions is an afternoon of AI. The value has moved to which one is right — and this section makes that judgment visible.
06Information design

Five decision views, each with one clear hero, and three layers of depth — glance → expand → source. You see the point first; you can open the detail; you can always trace the number.

NewsAI daily summary
ProductsProduct history & performance
MarketMarket comparison + full-year plan (TW/CN)
Store dailyCustomer pain points / try-on heat
SocialWhat content actually works
Two contexts, one source: a dark mode for daily monitoring, and a white "meeting mode" full-screen overlay for reporting — the same data dressed for the moment it's used in.
Dark · monitoring
Quarter vs last · illustrative
+12%YoY
Light · meeting mode
Quarter vs last · illustrative
+12%YoY
The same data, two modes side by side — dark for monitoring, light (light mode) for meetings; deliberately paired and labelled (illustrative)Dark / light comparison
LayerWhat it does · how the data connectsStack
SourcesERP snapshot, sales CSV, inventory, store POS and social APIs — raw, scattered, inconsistent.ERP · CSV · POS · APIs
PipelinePython scripts clean each source, validate it, auto-backup before every run, then write to the warehouse via row UPDATE — repeatable and safe.Python · gen_*.py
WarehouseA layered warehouse — dim/fact/mart/serving, 60 tables and 10 views — converging on the single trusted number for the whole company.BigQuery
Backend/data serves the data, /q runs live SQL; the whole service sits behind IAP, locked to the company domain — secure by default.Flask · Cloud Run · IAP
AI layerA daily summary reads the day's data and writes the takeaway; a plain-language Q&A turns natural language into a query and answers — so non-technical staff can ask the data anything.AI summary · NL→query
FrontendSingle-file UI, auto-binds data, inline fallback if a load fails, dark monitoring and white meeting modes.Vanilla · single-file
End-to-end data flow: sources → Python pipeline → BigQuery → Flask (Cloud Run + IAP) → AI layer → single-file frontend, built soloArchitecture at a glance
Social — what's actually workingSocial tab · last 30 days · GA-style analytics
ILLUSTRATIVE
Reach
48.2K+12%
Engagement
6.4%+0.8pt
Saves
1,310+18%
Profile taps
290−4%
Engagement · last 30 days
Day 18152230
Top contentReachEng. rateShare
Training reel · MAMBA12.4K8.1%
Race recap · Wuling9.8K7.2%
Product teaser7.1K5.4%
Athlete story5.3K6.0%
GA-style analytics surface (Social tab) — illustrative dummy figures, not real ops dataAnalytics surface
07Why it matters

From data model to pixel, this was a solo, full-stack build — data engineering, backend, deployment, security, AI and design system in one project. It's the clearest proof that I can take an ambiguous business problem and ship a working product around it, not just a screen.

Outcome: Shipped and live on Cloud Run behind IAP as the company's internal decision surface. It replaced the hand-uploaded daily report, and now pulls POS, CRM and other sources into one place — data entered once flows through to everywhere it's needed. The team actively uses its five decision views to review data, pull reports and run meetings. By an internal estimate, gathering and reconciling the numbers in one place cut that prep time by roughly 14%. Specific adoption figures stay internal.

Reflection: The value was never the charts — it was the judgment about what not to show, and the discipline to never fake a number. That's what makes data safe to decide on.

NEXT PROJECT
ZERO →
All work
© 2026 Alex HsuKPLUS — Operations HubTaipei, Taiwan