Projects

Each project is drawn as an architecture sheet. Select a component to read its note, or play the numbered flow. Nothing here connects to my live agents or servers.

Calibration automation and live dashboard

AI-built automation that runs calibration software offline and turns the results into a live, customer-ready dashboard.

Contract work through Insight Global as a Systems Integrator on a calibration team at Meta Reality Labs. Kept general on purpose. No screenshots, internal names, tools, data or process details.

In use since July 2026
Calibration automation and live dashboard. Calibration runs offline. AI automation collects and processes the results and uploads them to a SQL database that preserves the original data. An import job runs drift tracking, and the customer view combines the drift results with the database, using the charts and graphs most useful to the customer.
  • Running
  • In progress
  • Planned
  • 1Flow step

Select any component in the drawing to read its note.

Flow

Mock dashboard

Mock with made-up sample data. Not real results, and not the real dashboard.

Calibrations per dayTwo sample weeks against the built capacity of up to ~200 a day
50100150200Capacity, up to ~200
Sample day 1Day 14
Drift in one sample metricOffset from target over 30 sample runs, lower is better (arbitrary units)
0.20.40.6Change A: HelpfulChange B: Pause deployment
Sample run 1Run 30
Show the sample data as a table
Calibrations per day (sample)
DayCalibrations
1128
2142
3151
4137
5163
6171
7149
8158
9176
10165
11144
12181
13169
14174
Drift offset per run (sample)
RunOffsetChange
10.42
20.44
30.43
40.47
50.46
60.49
70.50
80.52
90.38Change A: Helpful
100.30
110.24
120.20
130.18
140.17
150.16
160.15
170.15
180.14
190.15
200.14
210.22Change B: Pause deployment
220.29
230.35
240.41
250.46
260.50
270.53
280.55
290.58
300.60
Problem
Calibration ran on station, blocking research studies in flight against long-running processes.
What I built

Built “automated systems using AI to run calibration software offline, culminate the data into a live dashboard and update/format that data for customer viewing”.

Processed results upload to a SQL database that preserves the original data. An import job runs the drift tracking, and the customer view combines the drift results with the database, using the charts and graphs that are most useful to the customer.

I also used agentic systems to trace which changes caused drift over time, and to judge whether each change helped or should pause a new deployment.

Stack
  • TypeScript web UI
  • SQL database
  • AI coding tools
Scale
Built to “handle up to ~200 calibrations daily” and “saves ~10mins per calibration”. 200 is capacity, not a daily total.
Status
In use since July 2026
What I learned
How to represent data to customers in a way that makes them care about what we do, and how to define complex data pipelines with automation built into each step, so processing runs quickly and continuously without stops.
Sheet
1 of 5
Drawing
Calibration automation
Status
In use since July 2026
Drawn by
David Kayne

Multi-agent orchestration layer

A layer, in progress, designed to let several AI agents from different companies hand work to each other.

In progress
Multi-agent orchestration layer. A private GitHub repository serves as a shared task board. Hermes is designed to read each task and delegate it to the most capable agent, who posts the result back. Automated pickup is not built yet.
  • Running
  • In progress
  • Planned
  • 1Flow step

Select any component in the drawing to read its note.

How it’s designed to work, one sample task

Problem
Re-explaining everything to every agent, with no real delegation based on which agent suits the task best.
What I’m building

A layer designed to let “Claude Duet, OpenAI Dots, Muse AI, and Hermes interact with eachother”, with Hermes as the orchestrator that delegates “to the most capable agent”.

A private GitHub repository acts as a shared task board the agents read and post to.

Today it can create and receive readable requests from multiple agents. There is no automated process yet for agents to pick up or respond to those requests.

Stack
  • Hermes Agent (orchestrator)
  • Claude Duet
  • OpenAI Dots
  • Muse AI
  • Private GitHub repository as the task board
Scale
None to state yet; it's in progress.
Status
In progress
What I learned
How to connect multiple complex out-of-the-box solutions using public APIs and GitHub, and how to use the tools available in otherwise locked-down environments to delegate out to multiple dispersed systems.
Sheet
2 of 5
Drawing
Multi-agent orchestration
Status
In progress
Drawn by
David Kayne

Hermes agents and the AI Dojo workspace

A self-hosted AI agent on AWS Lightsail that runs a shared AI workspace for a small Discord community.

Running since September 2026
Hermes agents and the AI Dojo workspace. Two Hermes agents. The one on AWS Lightsail works in AI Dojo, a Discord server, and uses Honcho as its memory provider (migrated from Hindsight). The one on Nous Cloud uses Hermes’ standard memory and skill management system and works on the ATLAS project.
  • Running
  • In progress
  • Planned

Select any component in the drawing to read its note.

Problem
I wanted to share production-ready harnesses and app integrations with my friends, to show them how these systems work and how they can help.
What I built

Two Hermes agents, one “On AWS Lightsail” and one “Through Nous Cloud”.

The Lightsail agent works in the Discord server AI Dojo, running “an AI workspace for me and my friends”. It uses Honcho as its memory provider, migrated from Hindsight.

The Nous Cloud agent uses Hermes’ standard memory and skill management system and works on the ATLAS project with me.

Stack
  • Nous Research's Hermes Agent, mainly running Sol 6.2
  • AWS Lightsail
  • Nous Cloud
  • Discord
  • Honcho (migrated from Hindsight)
  • Hermes memory and skill management
Scale
“6 people including myself”.
Status
Running since September 2026
What I learned
How to use agentic workspaces to collaborate and teach others, and how to use application statuses and key management to interact with one system in multiple ways and be more productive.
Sheet
3 of 5
Drawing
Hermes and AI Dojo
Status
Running since September 2026
Drawn by
David Kayne

ATLAS, conversational agent

An always-on conversational agent whose memory and context are managed by its own scripts and databases to keep costs down.

Ongoing since October 2026
ATLAS, conversational agent. Each Discord message goes to a classifier, which sends it either to an ATLAS LLM call on Luna 6 or to a Python script. Both paths pull from the SQLite database, and ATLAS replies in the Discord chat.
  • Running
  • In progress
  • Planned
  • 1Flow step

Select any component in the drawing to read its note.

Behind each message

Scripted chat

Written by hand for this page. Not a live connection to ATLAS, and not real messages.

  1. Member

    Atlas, we left the caravan at the river crossing last night. Where were we headed?

  2. ATLAS

    North, to the old observatory. You promised Mira you’d find the star chart before the first snow.

  3. Member

    Right, the chart. Do we still have the brass key?

  4. ATLAS

    Wren has it. She traded the lantern for it two nights ago, if you remember.

Problem
We wanted an AI conversation bot, and the project grew out of cost concerns.
What I built

An “AI configured role play bot with lots of supporting python scripts and databases to manage memory and conversation context for cost savings and always-on architecture”.

Each Discord message goes to a classifier, which sends it either to an ATLAS LLM call (Luna 6) or to a Python script. Both pull from the SQLite database.

A Hermes agent on Nous Cloud works on the ATLAS project with me.

Stack
  • Luna 6
  • Python
  • Message classifier
  • SQLite for memory and context
  • Discord
Scale
“Atlas has 12 people who role play with him in that server”.
Status
Ongoing since October 2026
What I learned
How to build a project for fun with friends: making something usable and seeing how it breaks, doing cost analysis on different features, and building in delegation to manage cost on repetitive tasks.
Sheet
4 of 5
Drawing
ATLAS
Status
Ongoing since October 2026
Drawn by
David Kayne

How this site is deployed

djkayne.com is itself a small cloud project, defined in Terraform and deployed by GitHub Actions without a single stored AWS key.

Live since November 2026
How this site is deployed. A push to main starts GitHub Actions, which type-checks and builds the site, signs in to AWS through an OIDC role that only trusts main, uploads to a private S3 bucket and clears the CloudFront cache. Visitors reach CloudFront through Route 53, over HTTPS with an ACM certificate.
  • Running
  • In progress
  • Planned
  • 1Flow step

Select any component in the drawing to read its note.

From git push to the live page

Problem
I wanted a portfolio that also shows the cloud and infrastructure skills I'm applying for.
What I built

The static Astro site lives in a private S3 bucket. Only CloudFront can read it, through origin access control. CloudFront serves it over HTTPS and HTTP/3 with AWS-managed security headers.

A CloudFront Function redirects www to the bare domain and maps clean URLs such as /projects/ to their index pages. The site's own 404 page handles anything missing.

A free ACM certificate is validated through Route 53, which also holds the A and AAAA alias records for both names.

All 19 AWS resources are defined in Terraform.

GitHub Actions type-checks and builds every pull request. Every push to main uploads the build to S3 and clears the CloudFront cache.

No AWS keys are stored anywhere. The workflow signs in with GitHub's short-lived OIDC token. The IAM role trusts only the production environment on main, matched on GitHub's immutable repository and owner IDs, and can only write one bucket and clear one distribution's cache.

Stack
  • Astro
  • TypeScript
  • Amazon S3
  • CloudFront and CloudFront Functions
  • AWS Certificate Manager
  • Route 53
  • IAM with GitHub OIDC
  • GitHub Actions
  • Terraform
Scale
Built for CloudFront's $0 flat-rate plan, on a domain I already own. A deploy takes about 40 seconds from push to live.
Status
Live since November 2026
Sheet
5 of 5
Drawing
This site's deployment
Status
Live since November 2026
Drawn by
David Kayne