Duncan AndersonIndependent systems
Back to selected work

Systems library / Selected operating workflows

Agents that do real work—and know where to stop.

A technical index of the AI agents, data workflows and controlled automations I have built across a live SaaS product, enterprise operations and client delivery.

Flagship workflow / Content and distribution

Daily Public Edge turns product evidence into finished short-form content.

This workflow changed content production from a blank-page scramble into a repeatable operating system. It finds a defensible story, gives me three sharply different editorial choices, and carries the selected idea through a static asset or an upload-ready founder video—with sources, captions, platform copy and approval receipts attached.

What changed

I can move from verified product evidence to a credible piece of public content while keeping the claim, visual and final publishing decision traceable. The system handles the research and production burden; I keep the editorial judgment and appear on camera.

A complete video package includes
  1. 01A 20–30 second, 45–75 word script and sentence-by-sentence iPhone recording plan
  2. 02Two to four product captures or data visuals tied to exact spoken lines
  3. 03Ordered portrait clips, corrected captions, conservative audio treatment and 1080×1920 export
  4. 04Cover image, Instagram and X copy, source note, checksum and verification receipt

Execution path

  1. 01Evidence

    Check current internal data, product state and official context.

  2. 02Choose

    Present exactly A / B / C: timely, original and product proof.

  3. 03Build

    Create the smallest truthful static asset or founder-led video.

  4. 04Verify

    Check sources, freshness, privacy, copy and final media export.

  5. 05Approve

    Stop on the exact package checksum before any public post.

  6. 06Learn

    Log the URL and measure real funnel outcomes after publication.

Selected systems / Technical index

The rest of the operating stack.

Open any system for its inputs, output contract, safety boundary and implementation stack. The business outcome stays visible; the engineering detail is one level deeper.

01Content productionDaily Public EdgeTurns current product evidence into a complete founder-led post or short-form video package without starting from a blank page.Operating workflowInspect system
Inputs

Current sports context, internal product evidence, recent story ledger, audience and channel goal.

System output

A verified story, visual or product capture, recording plan, edited vertical video, captions, cover, platform copy and source receipt.

Safety boundary

It cannot invent a result, reuse a stale claim or publish without approval of the exact final package.

Implementation
  • Python
  • SQL
  • Product data
  • Remotion
  • FFmpeg
  • Editorial ledger
02Business operationsProduction Operations WorkersKeeps data, model, growth and business-priority checks in one review system with a recorded next action.Internal production systemInspect system
Inputs

Scheduled checks, manual requests, provider evidence, application health and current operating priorities.

System output

A finding, evidence state, owner, proposed next step and durable review status.

Safety boundary

Workers can inspect and recommend; protected production changes remain approval-gated.

Implementation
  • Python
  • TypeScript
  • PostgreSQL
  • Supabase
  • Scheduled jobs
  • Approval gates
03Research automationEvidence-Backed Story MinerMaintains a ranked backlog of data stories so content starts from an answerable question and reproducible evidence.Research workflowInspect system
Inputs

Historical sports data, market records, settled results, prior story angles and an explicit research question.

System output

A ranked candidate with a query receipt, finding, caveats and a clear editorial handoff.

Safety boundary

Exploratory findings stay labelled as exploratory and negative results remain valid outcomes.

Implementation
  • Python
  • PostgreSQL
  • Supabase
  • Evidence ledger
  • Reproducible SQL
04Product reliabilityOdds and EV Integrity WatchDetermines whether customer-facing market data is current and publishable before the product or content relies on it.Read-only controlInspect system
Inputs

Quote age, provider coverage, publishability gates, board rows, scheduler lag and current production errors.

System output

A Green, Yellow, Red or Unavailable decision with the exact evidence and failure layer.

Safety boundary

The watch does not refresh data, change gates or call a degraded board healthy.

Implementation
  • Python
  • PostgreSQL
  • Sentry
  • Provider APIs
  • Freshness SLAs
05Machine learning operationsModel Champion PipelineMoves model work from an experiment to a reviewable candidate with reproducible evaluation and release gates.Controlled ML workflowInspect system
Inputs

Point-in-time features, model specification, untouched test data, benchmark and operational constraints.

System output

Versioned artifacts, evaluation evidence, calibration checks, promotion recommendation and monitoring plan.

Safety boundary

A promising experiment cannot promote itself; failed release evidence keeps the model out of production.

Implementation
  • Python
  • LightGBM
  • Feature engineering
  • Backtesting
  • Calibration
  • Monitoring
06Revenue operationsTrial and Billing RescueSeparates product friction and billing failures from users already covered by lifecycle automation.Exception workflowInspect system
Inputs

Checkout intent, trial activity, cancellation context, payment state and prior lifecycle messages.

System output

An automation-covered, automation-gap, product-friction, support-needed or approved-exception classification.

Safety boundary

No duplicate outreach, guessed billing state, unapproved discount or automatic customer contact.

Implementation
  • Stripe
  • RevenueCat
  • Postmark
  • PostgreSQL
  • Lifecycle events
07Enterprise AI operationsDispute DefenderPrepared reason-specific chargeback cases from booking, payment and customer-contact evidence for human review.Production enterprise systemInspect system
Inputs

Dispute reason, booking records, transaction history and customer communications.

System output

A traceable, case-specific evidence and response package instead of a generic template.

Safety boundary

The system organized source records; a reviewer retained the final response decision.

Implementation
  • Python
  • Machine learning
  • REST APIs
  • Data pipelines
  • Human review
08Client workflow automationDocument-to-Draft PublisherTurns approved Google Docs into structured, editable WordPress drafts without rebuilding each article by hand.Completed client phaseInspect system
Inputs

Approved document, content structure, SEO fields and reviewed image direction.

System output

Clean article content, editable metadata, reviewed imagery and an Elementor draft.

Safety boundary

The workflow creates a draft; clinical, editorial and publishing decisions remain human.

Implementation
  • Google Docs
  • WordPress REST API
  • Elementor
  • Yoast SEO
  • Next.js

Shared execution contract

Evidence before action. Approval before consequence. Receipts after execution.

  1. 01Evidence

    Use the current system of record, timestamp the claim and surface unavailable inputs.

  2. 02Action

    Give each run a bounded objective, named owner, explicit output and failure state.

  3. 03Approval

    Stop before deployment, publication, outreach, billing or another protected action.

  4. 04Receipt

    Keep the artifact, decision, checksum or authoritative state that proves what happened.

Technical fit

Need an engineer who can own the workflow, not just the model call?

I work across product, data, APIs, AI behavior, reliability and operator handoff. Bring me the workflow and the consequence of getting it wrong.

Book a 30-minute fit call Read the AI Operations case study