Project showcase

Harbormaster and Spatial Atlas

These two narrated talks each walk through one project I built. Part 1 covers Harbormaster, a real-time maritime anomaly-detection platform on AWS. Part 2 covers Spatial Atlas, a compute-grounded spatial reasoning agent.

By Arun Sharma

The shared question

The same question in both projects

Harbormaster · Part 1 of 2

Does a maritime alert stay attributable, reviewable, and recoverable?

Spatial Atlas · Part 2 of 2

Is a spatial answer supported by explicit evidence and computation?

Both projects share one research question about evidence. In both projects, I want each output backed by evidence that someone can check.

Part 1 of 2

Harbormaster

A solo build of a real-time maritime anomaly-detection platform on AWS

Title card for the Harbormaster talk, Project showcase 1 of 2

I built Harbormaster solo, and its core idea is a fast physics check before the heavier model. The talk walks through the two-plane architecture and replay-safe change data capture. It also covers the four promotion gates that every candidate model must pass in order. It then shows the audit that caught a scale-to-zero bug before I ran apply.

  • 34,123records in 15 minutes
  • 35,999HTTP 200 responses out of 36,000 in a one-hour soak
  • 1.89 sfrom scale request to first HTTP 200
  • 19findings in the SageMaker Terraform audit
Chapters for Harbormaster15
  1. 0:00Harbormaster, a solo build on AWS
  2. 0:24Different failures, different checks
  3. 1:05Five capability gaps, named before code
  4. 1:30A fast physics check before the model
  5. 2:10Two planes, research trains and AWS serves
  6. 2:47Teardown by apply, never destroy
  7. 3:19Exactly-once results from at-least-once
  8. 4:0834,123 records, then lag drained to 0
  9. 4:47One-hour soak, exactly one 503
  10. 5:22Gates that distrust the model
  11. 6:08The audit caught a scale-to-zero bug
  12. 6:51One human, agents as tooling
  13. 7:28What I learned and what I would change
  14. 8:19Fakes fail physics before they fail classifiers
  15. 8:35Thank you

Part 2 of 2

Spatial Atlas

A compute-grounded spatial reasoning agent

Title card for the Spatial Atlas talk, Project showcase 2 of 2

I built Spatial Atlas for the Berkeley RDI AgentX-AgentBeats competition. It computes spatial facts from a typed scene graph before a language model writes its answer. The model therefore does not have to guess distances from pixels. An opt-in strict metric bridge rebuilds geometry from masks and depth with NVIDIA's SpatialClaw tools. The bridge fails closed when its evidence checks do not pass, so it returns an explicit unavailable result.

  • 4paths ran end to end with zero retries
  • 8label-free rows per path, with no correct answers in view
  • 323tests in the regression suite
  • 92.09%coverage of the perception module
Chapters for Spatial Atlas15
  1. 0:00Spatial Atlas and the research question
  2. 0:22Guessing distances from pixels
  3. 0:57Compute first, then let the model write
  4. 1:35Spatial Atlas, an Agent-to-Agent server
  5. 2:09How a question becomes an answer
  6. 2:59The scene becomes a typed scene graph
  7. 3:37Geometry rebuilt from masks and depth
  8. 4:32If checks fail, the bridge fails closed
  9. 4:59The whole system on one map
  10. 5:34The ML path runs no code by default
  11. 6:10All four paths ran end to end
  12. 6:58The repository carries its own tests
  13. 7:27The design, in five choices
  14. 8:22The same question in both projects
  15. 8:53Thank you and links

Series order

Watch in order

Part 1 on Harbormaster comes first, and Part 2 on Spatial Atlas follows it. The second-to-last slide of the Spatial Atlas talk sets the questions from both projects side by side.