STRANDED COMPUTE
Distributed AI on the
computers already on your desk
A coordinator routes work to local Ollama workers,
then makes ownership and runtime visible.
HACKATHON DEMO · ABEL · KEVIN · RONALD
ONE COORDINATOR
Coordinator API
Abel
Kevin
Ronald
local workers
WHY THIS EXISTS
Idle compute is useful only when work can find it
The demo turns spare Mac capacity into a visible, coordinator-managed worker pool.
STRANDED COMPUTE / 02
BEFORE
One laptop becomes the bottleneck
1
Capacity sits idle
A second Mac has RAM, GPU time, or a loaded model that no job can reach.
2
Ownership is hard to see
A queue tells you that work is waiting, not which computer is doing it.
AFTER
A shared pool with a clear owner
The coordinator knows which worker is online,
which model it advertises, and what it is running.
SELECT → CLAIM → MEASURE
Local inference stays on each participating Mac.
SYSTEM
The coordinator is the control plane
Browsers submit intent. Workers execute locally. The database keeps the run auditable.
STRANDED COMPUTE / 03
USER BROWSER
Choose task
Choose model
Choose active worker
COORDINATOR API
Queue and scheduler
eligibility · leases · retries
worker heartbeats · activity metrics
WORKER POOL
Abel
Gemma 12B
Kevin
Qwen Coder
Heartbeat
Ollama
Each worker reports hardware, models, capacity, and task status.
POSTGRESQL
Jobs · tasks · results · telemetry
EXECUTION
Every job has a visible path
The same flow supports summaries, document Q&A, extraction, and coding assistance.
STRANDED COMPUTE / 04
01
Choose
Task type, model, and active Mac
02
Queue
Coordinator creates one or more tasks
03
Claim
Heartbeat worker takes a compatible lease
04
Generate
Ollama runs the selected local model
05
Persist
Result, timing, and owner become visible
Lease expiry and retry handling keep stale workers from blocking the queue.
USER CONTROL
The submission form makes the destination explicit
The user chooses a real model and an active machine before the coordinator accepts the job.
STRANDED COMPUTE / 05
CHOOSE A MODEL
Only models advertised
by active workers appear.
gemma3:12b
qwen2.5-coder:3b
No hidden coordinator default.
The model is part of the job contract.
CHOOSE WHERE WORK RUNS
Pick an active Mac
before submitting.
1
Availability
Only active workers are shown in the machine list.
2
Capacity
RAM, GPU, model, and current task status are visible.
OBSERVABILITY
The result includes who did the work
A demo run can show distribution instead of asking the audience to trust the queue.
STRANDED COMPUTE / 06
DEMO SNAPSHOT
6 tasks completed across 2 Macs
The dashboard attributes each task to its worker and records execution timing.
5
1
Abel-Mac
Kevin-Mac
TASK DETAILS
Computer
Abel-Mac / Kevin-Mac
Model
Gemma 3 12B / Qwen Coder
Task
Summarization
Queue time
Time waiting before claim
Execution
Worker-reported runtime
Attempts
Retries and final owner
6 / 6 COMPLETE
LIVE DEMO
A short path from prompt to proof
The audience can follow the request, assignment, and result without leaving the dashboard.
STRANDED COMPUTE / 07
01
Connect
Open the HTTP dashboard and enter the coordinator token.
02
Select
Choose a task, a concrete model, and an active Mac.
03
Submit
The coordinator creates the queued task and lease.
04
Watch
Telemetry shows the worker, status, queue, and runtime.
05
Explain
Click Recent task activity for the full task record.
The demo stays local to the participating Macs while the coordinator keeps one shared view.
WHAT THE DEMO PROVES
One coordinator.
Several Macs.
Measurable work.
The system turns local models and spare capacity into a small, inspectable compute network.
DEMO BOUNDARIES
1
Local models
Ollama runs inference on each worker Mac.
2
Coordinator control
The API owns queue state, leases, retries, and telemetry.
3
Narrow network scope
LAN demo with token access. No payments, public discovery, or NAT traversal.
NEXT STEP: account identity and provider approvals