Duncan AndersonIndependent systems
Back to the consulting site

Selected work / Case studies

What I built, what changed, and what I can prove.

These are working systems, not capability demos. Each example names the recurring problem, my role, the verifiable result, and the decision that still belongs to a person.

Client example / Healthcare practice

A private leadership dashboard for a clinic owner.

The problemThe owner had Jane reports and bank exports, but no dependable weekly view of revenue, appointments, therapist capacity, expenses, or which figures were still incomplete.

What I builtI built a guided refresh process and private dashboard. It recognizes the agreed Jane reports, can include bank files when financial reporting needs them, removes direct client details, and carries source warnings into the owner’s review instead of filling gaps with guesses.

What can be verifiedVerified here: the private dashboard and refresh process work end to end; current reporting periods and incomplete sources are named; raw uploads and direct client identifiers are not retained. Client metrics and screenshots remain private.

What stays humanThe dashboard prepares the review. The owner and accountant still decide how to interpret incomplete periods, cash position, reserves, and any figure that needs financial judgment.

Four systems / Four kinds of proof

A live subscription product, its business-monitoring system, a confidential enterprise tool, and a client publishing process.

  1. 01The LineupLive product
  2. 02The Lineup AI OperationsProduction operations system
  3. 03Dispute DefenderProduction enterprise system
  4. 04Trauma Therapy Group PublisherClient delivery
How the system worksIllustrated example

The Lineup

Live customer product

Live

01

Live inputs

Current market context

02

Model view

Projection and confidence

03

Recorded result

Automatically graded

One decision, full context

ML

The context stays attached to the decision through the final result.

What stays attached

PriceTimestamped
Model contextAttached
Final resultRecorded

Illustration based on the live product · public results remain available on The Lineup

01 / Sports analytics

Live product · Founder, product owner, solo builder

The Lineup

A live sports product that connects projections, market prices, subscriptions, and recorded results.

What changed

Customers can compare a projection with current prices, make their own decision, and return later to see the recorded result.

What you can verify
  1. 01Public web product
  2. 02Published iOS application
  3. 03Automated grading and public result history
What stays human

A projection is never shown on its own. The current price, timestamp, model context, and final result stay attached. The customer still decides whether any opportunity is worth acting on.

How the system worksIllustrated example

The Lineup

Business monitoring

Four workers
Data and pricing
Models and forecasts
User growth
Business priorities

Latest finding

Freshness issue needs review

01Issue recorded
02Next step proposed
03Production change paused
Duncan must approve the change

Illustration based on the private review dashboard · no production data shown

02 / Business monitoring

Production operations system · Founder, system designer, and operator

The Lineup AI Operations

Four specialized AI workers monitor the business—but they cannot quietly change production.

What changed

The recurring checks now happen on a schedule, issues are kept in one review queue, and production changes cannot happen without explicit approval.

What you can verify
  1. 01Four named workers in one review system
  2. 02Scheduled and manual checks with a recorded finding and next action
  3. 03Approval and reliability checks keep incomplete or blocked work visible
What stays human

The workers can inspect and recommend; they cannot quietly deploy, suppress a model, contact a customer, or make another protected production change. I review and approve those decisions. The system also keeps a check separate from a verified fix.

How the system worksIllustrated example

Case 02491

Service provided

Review
Booking record2 records
Payment history3 records
Customer communication4 records

Evidence package

Ready for operator review

Sources linked

8 / 8

Missing evidence

None detected

A person approves the final response. Source records remain attached to the case.

Representative process · identifying data removed

03 / Travel operations

Production enterprise system · Data scientist and system builder while employed

Dispute Defender

Turn a high-volume dispute queue into evidence-backed responses.

What changed

Reviewers received a case-specific set of source records instead of rebuilding every dispute by hand or relying on a generic response.

What you can verify
  1. 01Built for high-volume operational use
  2. 02Evidence remained traceable to source records
  3. 03Production enterprise deployment
What stays human

The system organized existing records; it did not invent facts or make the final judgment. A reviewer approved the response. The public illustration is representative because the company and customer records are confidential.

How the system worksIllustrated example

01

Google Doc

Approved source

02

Review

SEO + image

03

WordPress

Draft only

Article draft

When anxiety does not feel temporary

Draft
FormattingReady
SEO fieldsNeeds approval
Featured imageSelected
Publication remains a human action inside WordPress.

Representative process · client content anonymized

04 / Healthcare content operations

Client delivery · Process designer and implementation consultant

Trauma Therapy Group Publisher

Move a reviewed Google Doc into WordPress without rebuilding the article by hand.

What changed

The team can move approved copy into a structured WordPress draft without rebuilding the article by hand.

What you can verify
  1. 01Completed client phase
  2. 02End-to-end document-to-draft process
  3. 03Human review retained before publication
What stays human

The automation prepares and organizes the draft. It does not make the final clinical or publishing decision, and it does not auto-publish content.

Bring the recurring work

Is a recurring process still harder to trust than it should be?

Bring one recurring process and the point where it keeps breaking. We will decide whether it has a clear enough finish line for a sprint.