Skip to contentSkip to contact
Dennis KhorScreenrights
Recruiter mode
CareerAug 2024 to nowSydney, hybridRole 01 of 02

Screenrights

Building the portals where members manage their royalties.

My simplified illustration: claims arrive from many sources, pass through the member portals and staff tools and leave as royalties to many rights holders.

OverviewAug 2024 to nowSydney, hybrid

Six chapters at Screenrights

Stack
  • Python
  • Slack Bolt
  • Amazon Bedrock
  • Claude
  • AWS ECS Fargate
  • Terraform
  • AWS Cognito
  • Auth0
  • AWS CDK
  • NestJS
  • React
  • TypeScript
  • Retool
  • AI claim analysis
  • AWS Lambda
  • Amazon DocumentDB
  • Amazon EventBridge
  • Amazon SQS
  • .NET
  • Vite
  • Cloudflare Workers
  • Durable Objects
  • WebSockets
  • three.js
  • Vitest
  • 04 Replaced a manual membership process6transactional email templates, each sent on the right status change
  • 05 Kept member tasks in sync2 of 3duplicate or out-of-order events never create duplicate tasks; event consumers built
  • 06 Shipped a real-time planning-poker app in a daySame dayfrom the first design note to a live tool, then four feature rounds in three days

What I bring

Four strengths, each linked to the chapters that prove it.

  1. Prototype fast, then ship what earns it

    A working prototype in front of stakeholders within days, real feedback, then production quality where it pays off. I build for the business, not for myself.

  2. AI products that are safe to ship

    Tool-calling agents with cited answers, a person in the loop and cost controls.

  3. Full-stack delivery with no downtime

    React, NestJS and Python on AWS, shipped to production; migration plans, feature flags and rollback paths that let a sign-in switch-over land with no maintenance window.

  4. Owning the work end to end

    Strategy docs, decision records, UAT, sign-off for my area, runbooks and handovers; I pick up unowned work and finish it.

01 / 06Proposed growing it from an early prototype, then built most of it

Grew an internal AI assistant to five live sources, with cited answers

When
Aug 2025 to now
Team
2 engineers
Its own worldEnter Neo
How a question flows

Press play to walk through it in 5 steps, or use next.

Parts marked with a plus have a note. Hover, focus or tap one.

Tools only read

Nothing is filed or published unless a person asks, and pages are clearly marked as AI-generated.

Slack

A Slack assistant answers staff questions from the team's own internal sources.

Assistant service

A provider-neutral model layer made the move to Amazon Bedrock a config flip that could roll back with no redeploy.

Model running a bounded tool loop

Swapped the vector index for read-only tools that search each system live: more tokens per question, but always current and nothing to re-index.

Ticket and page drafting

Built ticket and page drafting with a person in the loop.

Answer with sources

It shows where every answer came from, across 5 live sources, up from 2 static document sets.

A question comes in from Slack, as a mention, a direct message or in the assistant sidebar, and goes to the assistant service, which handles the event and keeps thread memory. The model runs a bounded tool loop that searches the team's internal sources live, with read-only tools. It returns an answer with its sources and a confidence footer; a thumbs-down asks which source was wrong, with one button per source, and posts a comment on that source page or ticket. Tools only read, a person asks before anything is written, and the loop has a step limit; ticket and page drafting plugs into the same Slack app and model layer.

Step by step:

  1. A Slack assistant answers staff questions from the team's own internal sources.
  2. Read-only tools search each system live: more tokens per question, but always current and nothing to re-index.
  3. It shows where every answer came from.
  4. A provider-neutral model layer made the move to Amazon Bedrock a config flip that could roll back with no redeploy.
  5. Ticket and page drafting with a person in the loop. Nothing is filed or published unless a person asks.

Notes:

  • Tools only read: Nothing is filed or published unless a person asks, and pages are clearly marked as AI-generated.
  • Slack: A Slack assistant answers staff questions from the team's own internal sources.
  • Assistant service: A provider-neutral model layer made the move to Amazon Bedrock a config flip that could roll back with no redeploy.
  • Model running a bounded tool loop: Swapped the vector index for read-only tools that search each system live: more tokens per question, but always current and nothing to re-index.
  • Ticket and page drafting: Built ticket and page drafting with a person in the loop.
  • Answer with sources: It shows where every answer came from, across 5 live sources, up from 2 static document sets.

Connections:

  • Slack to Assistant service
  • Assistant service to Model running a bounded tool loop
  • Model running a bounded tool loop to Internal sources
  • Model running a bounded tool loop to Answer with sources
  • Answer with sources to Thumbs-down
  • Thumbs-down to Which source was wrong?
  • Which source was wrong? to Comment on the source
  • Assistant service to Ticket and page drafting

A Slack assistant that answers staff questions from the team's own internal sources and shows where every answer came from. I replaced its vector-search pipeline with live tool-calling search and led the assistant side of its move onto Claude on Amazon Bedrock. I also built ticket and page drafting with a person in the loop.

How I did it
  1. Swapped the vector index for read-only tools that search each system live: more tokens per question, but always current and nothing to re-index.
  2. A provider-neutral model layer made the move to Amazon Bedrock a config flip that could roll back with no redeploy.
  3. Nothing is filed or published unless a person asks, and pages are clearly marked as AI-generated.
  • Python
  • Slack Bolt
  • Amazon Bedrock
  • Claude
  • AWS ECS Fargate
  • Terraform

Outcomes

  • 5 live sources, up from 2 static document sets, with screenshot input and conversation memory
  • Every deploy pinned to a commit and reversible in one step

02 / 06Led the migration and built the new-portal side

Led the member sign-in move from Auth0 to AWS Cognito with no maintenance window

When
2026
Team
4 to 5 engineers
How members moved to the new sign-in

Press play to walk through it in 4 steps, or use next.

Parts marked with a plus have a note. Hover, focus or tap one.

Member

Members moved across as they signed in, with no bulk password reset and no maintenance window.

New member portal

Led the migration and built the new-portal side.

Switch per app

Each app could switch back to the old sign-in on its own during the move.

Old identity provider

Each app could switch back to the old sign-in on its own during the move.

Migration step

Chose a move-on-sign-in migration over a bulk import, so nobody was sent a bulk password reset.

A member signs in through one of three apps, a new member portal, an older member portal and a staff tool. A per-app switch sends each to the new sign-in, or back to the old provider as a rollback. A member not yet moved is moved across on that first sign-in, so nobody was sent a bulk password reset.

Step by step:

  1. Member sign-in for the member portals and staff tool moves from Auth0 to AWS Cognito, without taking the portals offline.
  2. Members moved across as they signed in.
  3. Chose a move-on-sign-in migration over a bulk import, so nobody was sent a bulk password reset.
  4. Each app could switch back to the old sign-in on its own during the move.

Notes:

  • Member: Members moved across as they signed in, with no bulk password reset and no maintenance window.
  • New member portal: Led the migration and built the new-portal side.
  • Switch per app: Each app could switch back to the old sign-in on its own during the move.
  • Old identity provider: Each app could switch back to the old sign-in on its own during the move.
  • Migration step: Chose a move-on-sign-in migration over a bulk import, so nobody was sent a bulk password reset.

Connections:

  • Member to New member portal
  • Member to Older member portal
  • Member to Staff tool
  • New member portal to Switch per app
  • Older member portal to Switch per app
  • Staff tool to Switch per app
  • Switch per app to New sign-in
  • Switch per app to Old identity provider
  • New sign-in to Account exists?
  • Account exists? to Signed in
  • Account exists? to Migration step
  • Old identity provider to Migration step
  • Migration step to Create account: yes
  • Create account to Signed in: signed in

Moved member sign-in for the organisation's member portals and staff tool from Auth0 to AWS Cognito without a bulk password reset or taking the portals offline. Members moved across as they signed in.

How I did it
  1. Chose a move-on-sign-in migration over a bulk import, so nobody was sent a bulk password reset.
  2. Each app could switch back to the old sign-in on its own during the move.
  3. Wrote the migration strategy, the rollout and rollback plan and the UAT plan, then owned sign-off for the new portal at the go or no-go.
  • AWS Cognito
  • Auth0
  • AWS CDK
  • NestJS
  • React
  • TypeScript

Outcomes

  • Members moved across as they signed in, without a bulk password reset
  • No maintenance window and no rollback at the switch-over

03 / 06Wrote the design; built and demoed the prototype

First place at the company hackathon, with an AI prototype

When
Sep to Oct 2025
Team
3, with me as the only engineer
Automating the admin of competing claims

Press play to walk through it in 5 steps, or use next.

Parts marked with a plus have a note. Hover, focus or tap one.

Intake

Intake used simple keyword rules.

Case dashboard

Breadth over polish: working pieces across the whole workflow in two days.

AI claim analysis

AI had one job, claim analysis with its reasoning on screen.

Staff approval queue

Automated follow-ups wait for a person to approve them.

Inside an outline marked sample data, not connected to live systems, inbound request emails go to intake, which uses keyword rules to find the title and claimant, then to a case dashboard, a phase launcher, a deadline tracker and templated mail-outs. From the dashboard a branch runs to AI claim analysis, which shows its reasoning, then a staff approval queue and follow-up actions. The deadline tracker and the approval queue both feed chat notifications.

Step by step:

  1. A prototype that automates the admin of competing claims, when two rights holders register the same title, starting with email intake.
  2. A case dashboard, phase and deadline tracking, and templated mail-outs.
  3. AI had one job, claim analysis with its reasoning on screen; intake used simple keyword rules.
  4. Automated follow-ups wait for a person to approve them.
  5. Breadth over polish: a working end-to-end prototype in a two-day build, on sample data.

Notes:

  • Intake: Intake used simple keyword rules.
  • Case dashboard: Breadth over polish: working pieces across the whole workflow in two days.
  • AI claim analysis: AI had one job, claim analysis with its reasoning on screen.
  • Staff approval queue: Automated follow-ups wait for a person to approve them.

Connections:

  • Inbound request emails to Intake
  • Intake to Case dashboard
  • Case dashboard to Phase launcher
  • Phase launcher to Deadline tracker
  • Deadline tracker to Templated mail-outs
  • Case dashboard to AI claim analysis
  • AI claim analysis to Staff approval queue
  • Staff approval queue to Follow-up actions
  • Deadline tracker to Chat notifications
  • Staff approval queue to Chat notifications

A prototype that automates the admin of competing claims, when two rights holders register the same title: email intake, a case dashboard, phase and deadline tracking, templated mail-outs and AI claim analysis that shows its reasoning, with follow-ups waiting for staff approval.

How I did it
  1. AI had one job, claim analysis with its reasoning on screen; intake used simple keyword rules.
  2. Breadth over polish: working pieces across the whole workflow in two days.
  3. Automated follow-ups wait for a person to approve them.
  • Retool
  • AI claim analysis

Outcomes

  • First place out of seven teams, by anonymous company-wide staff vote
  • A working end-to-end prototype in a two-day build, on sample data

04 / 06Main app-side developer

Replaced a manual membership process with one guided workflow

When
May 2025 to Jan 2026
Team
About 10 to 12 across app and data teams
Membership status lifecycle

Press play to walk through it in 5 steps, or use next.

Parts marked with a plus have a note. Hover, focus or tap one.

Non-member

Replaced application kits, PDF records, follow-up emails and hand-copied records with one guided workflow.

In review

Staff review, request changes, reject or approve under separate roles, and every step sends its own notification.

Pending approval

Reviewing and approving are separate roles: a reviewer and a separate approver.

Current member

6 transactional email templates, each sent on the right status change.

Application sent

A staff-started path, alongside applications the applicant starts.

A state diagram of a membership application. A non-member applies and goes to In review, or staff start the application, which goes to Application sent and then, once the applicant accepts the agreement, to In review. A reviewer moves it to Pending approval and an approver to Current member; the reviewer or the approver can ask for changes, which moves it to Awaiting update until the applicant updates it, or reject it, which ends in Rejected. Staff can later move a current member to Resigned or Inactive.

Step by step:

  1. Applicants apply and follow their status.
  2. The status model also has a staff-started path.
  3. Staff review and request changes.
  4. They reject or approve under separate roles, a reviewer and a separate approver.
  5. A four-state model grew into about ten statuses, and every step sends its own notification.

Notes:

  • Non-member: Replaced application kits, PDF records, follow-up emails and hand-copied records with one guided workflow.
  • In review: Staff review, request changes, reject or approve under separate roles, and every step sends its own notification.
  • Pending approval: Reviewing and approving are separate roles: a reviewer and a separate approver.
  • Current member: 6 transactional email templates, each sent on the right status change.
  • Application sent: A staff-started path, alongside applications the applicant starts.

Connections:

  • Non-member to In review: Applicant
  • Non-member to Application sent
  • Application sent to Applicant accepts agreement
  • Applicant accepts agreement to In review
  • In review to Pending approval: Reviewer
  • Pending approval to Current member: Approver
  • In review to Awaiting update
  • Pending approval to Awaiting update
  • In review to Rejected
  • Pending approval to Rejected
  • Current member to Resigned
  • Current member to Inactive

Replaced application kits, PDF records, follow-up emails and hand-copied records with one guided workflow: applicants apply and follow their status, staff review, request changes, reject or approve under separate roles, and every step sends its own notification.

How I did it
  1. Grew a four-state model into about ten statuses, with a reviewer and a separate approver and a staff-started path.
  2. An EventBridge-triggered Lambda keeps every linked workspace in step.
  3. Kept the legacy statuses working so the old and new systems could run side by side.
  • NestJS
  • AWS Lambda
  • Amazon DocumentDB
  • Amazon EventBridge
  • React
  • Retool
  • TypeScript

Outcomes

  • Shipped to production in late 2025, and member services ran applications through it from launch day
  • 6 transactional email templates, each sent on the right status change
  • The new platform became the system of record, with the legacy database kept in step

05 / 06Built two of the three event consumers

Kept member tasks in sync with a legacy claims system, duplicate-proof

When
Feb to Sep 2026
Team
4 to 5 engineers; my slice solo
How member tasks stay in sync

Press play to walk through it in 6 steps, or use next.

Parts marked with a plus have a note. Hover, focus or tap one.

Duplicate or out-of-order event

An update that arrives before its create is treated as a create.

Event bus

Events can arrive twice, late or out of order.

Serverless consumer

Event-driven consumers turn claims work in the legacy system into a to-do list in the member portal.

Create or update

Idempotent create-or-update writes, keyed on title and workspace.

Member portal

No duplicate tasks, even when events arrive twice, late or out of order.

Explicitly active

Only an explicitly active status reopens a task. A final status completes it, and deleted stays deleted.

A legacy claims system (.NET) sends created, updated and deleted events to an event bus with one rule per consumer, then to a serverless consumer. The consumer creates or updates each task, keyed on title and workspace, in a document store of tasks and titles that feeds the task list and count badges in the member portal. A duplicate or out-of-order event loops back into the consumer and never creates a duplicate task. A status ladder is checked in order: deleted stays deleted, a final status completes with the time recorded once, an explicitly active status reopens, and anything else refreshes details only.

Step by step:

  1. Claims work happens in a legacy .NET system, and its changes go out as events.
  2. Event-driven consumers pick them up, even when events arrive twice, late or out of order.
  3. Writes are idempotent: create or update, keyed on title and workspace.
  4. The result is a to-do list in the member portal with no duplicate tasks.
  5. Events may arrive twice, late or out of order. An update that arrives before its create is treated as a create.
  6. A task-lifecycle state machine: a final status completes, only an explicitly active status reopens, and deleted stays deleted.

Notes:

  • Duplicate or out-of-order event: An update that arrives before its create is treated as a create.
  • Event bus: Events can arrive twice, late or out of order.
  • Serverless consumer: Event-driven consumers turn claims work in the legacy system into a to-do list in the member portal.
  • Create or update: Idempotent create-or-update writes, keyed on title and workspace.
  • Member portal: No duplicate tasks, even when events arrive twice, late or out of order.
  • Explicitly active: Only an explicitly active status reopens a task. A final status completes it, and deleted stays deleted.

Connections:

  • Legacy claims system to Event bus: created, updated or deleted
  • Event bus to Serverless consumer
  • Serverless consumer to Create or update
  • Create or update to Document store
  • Document store to Member portal
  • Event bus to Duplicate or out-of-order event
  • Duplicate or out-of-order event to Serverless consumer
  • Serverless consumer to Deleted
  • Deleted to Final status
  • Final status to Explicitly active
  • Explicitly active to Anything else

Event-driven consumers that turn claims work in a legacy .NET system into a to-do list in the member portal, built so events that arrive twice, late or out of order never create a duplicate task.

How I did it
  1. Idempotent create-or-update writes keyed on title and workspace; an update that arrives before its create is treated as a create.
  2. A task-lifecycle state machine: a final status completes, only an explicitly active status reopens, and deleted stays deleted.
  • TypeScript
  • NestJS
  • AWS Lambda
  • Amazon EventBridge
  • Amazon SQS
  • Amazon DocumentDB
  • AWS CDK
  • .NET

Outcomes

  • 2 of the 3 event consumers built; duplicate or out-of-order events never create duplicate tasks

06 / 06Built it end to end with an AI coding agent

Shipped a real-time planning-poker app in a day, on a free tier

When
Jul to Aug 2026
Team
Solo
One room, one home, one reveal gate

Press play to walk through it in 5 steps, or use next.

Parts marked with a plus have a note. Hover, focus or tap one.

Player laptops

A real-time estimation app for my team's weekly planning: private rooms, live voting and chat.

Live connections + presence

One Durable Object per room gives each room's state exactly one home.

Game rules

The game rules are pure functions with unit tests.

Reveal gate

One server-side reveal gate enforces the no-peeking rule. Votes never leave the server before the reveal, checked on the wire with an observer connection.

Idle rooms

Idle rooms hibernate at no cost, so it fits Cloudflare's free tier.

3D table

The 3D table keeps ordinary HTML controls on top. Its graphics memory was cut by more than half on a high-density screen, with no visible change.

Player laptops and the meeting-room screen, in big-screen mode, load the page from static files at the edge, where no code runs. Each browser holds a two-way live connection to an edge Worker, which routes it to its room object: one per room, holding live connections and presence, the game rules as pure functions, and a reveal gate where votes stay until the reveal; idle rooms hibernate. On reveal, the room pushes the full room state to every screen. In each browser, HTML controls sit over a 3D table that loads on demand, with a CSS table when there is no WebGL.

Step by step:

  1. A real-time estimation app for my team's weekly planning: private rooms, live voting and chat.
  2. One Durable Object per room gives each room's state exactly one home, and idle rooms hibernate at no cost.
  3. One server-side reveal gate enforces the no-peeking rule; the game rules are pure functions with unit tests.
  4. Nobody can see a vote before the reveal, not even in the browser's developer tools: votes never leave the server until then.
  5. The 3D table keeps ordinary HTML controls on top, with a plain CSS table when there is no WebGL.

Notes:

  • Player laptops: A real-time estimation app for my team's weekly planning: private rooms, live voting and chat.
  • Live connections + presence: One Durable Object per room gives each room's state exactly one home.
  • Game rules: The game rules are pure functions with unit tests.
  • Reveal gate: One server-side reveal gate enforces the no-peeking rule. Votes never leave the server before the reveal, checked on the wire with an observer connection.
  • Idle rooms: Idle rooms hibernate at no cost, so it fits Cloudflare's free tier.
  • 3D table: The 3D table keeps ordinary HTML controls on top. Its graphics memory was cut by more than half on a high-density screen, with no visible change.

Connections:

  • Edge: static files to Player laptops: page load
  • Edge: static files to Meeting-room screen
  • Player laptops to Edge Worker
  • Meeting-room screen to Edge Worker
  • Edge Worker to Live connections + presence
  • Live connections + presence to Game rules
  • Game rules to Reveal gate
  • Reveal gate to Live connections + presence: on reveal: full room state
  • HTML controls to 3D table
  • 3D table to CSS table

A real-time estimation app for my team's weekly planning: private rooms, live voting and chat, and one rule that nobody can see a vote before the reveal, not even in the browser's developer tools. It fits Cloudflare's free tier.

How I did it
  1. One Durable Object per room gives each room's state exactly one home, and idle rooms hibernate at no cost.
  2. One server-side reveal gate enforces the no-peeking rule; the game rules are pure functions with unit tests.
  3. The 3D table keeps ordinary HTML controls on top, with a plain CSS table when there is no WebGL.
  • React
  • TypeScript
  • Vite
  • Cloudflare Workers
  • Durable Objects
  • WebSockets
  • three.js
  • Vitest

Outcomes

  • Same day from the first design note to a live tool, then four feature rounds in three days
  • Votes never leave the server before the reveal, checked on the wire with an observer connection
  • Cut the 3D view's graphics memory by more than half on a high-density screen, with no visible change

Skills shown here

Each group links to the chapters that show it.

Smaller wins

Code review
Several hundred reviews of my teammates' pull requests.
Roles and permissions as code
Replaced a manual permission process with one validated, version-controlled source of truth that CI syncs.
Test infrastructure
Introduced visual regression testing and wrote the team's end-to-end testing handbook.

A personal portfolio. Not an official Screenrights page and not endorsed by Screenrights. For membership, claims or payments, contact Screenrights directly.

Ask about fit

Paste a job description. You get a verdict, the evidence that matches it with citations, and what the sources do not show.

Open to senior AI and software engineering roles: on site or hybrid in Sydney, or remote for Australian and US teams.

Email meBook 15 minCV