AI Product Prototype

Atlas

An analyst triage tool for AI-detected objects in overhead imagery — built for the moment the model is good enough and the human becomes the bottleneck.

Scoped and shipped solo in 5 days · PRD → build → deploy → recorded walkthrough

Next.js 14TailwindYOLOv8mSupabaseVercelxView dataset
The Problem

The analyst is the bottleneck.

Atlas is an interface for an intelligence analyst to review, accept or reject, and dispatch AI-generated object detections on satellite or drone imagery. It optimises for time-to-dispatch under accuracy constraints — once the model is good enough, the product lives in how fast verification can happen without collapsing trust.

This is not a real defense product. It's a 5-day prototype built to demonstrate product thinking, AI-native shipping, and taste in workflow design — and it's honest about that scope: no trained custom model (YOLOv8 pretrained, run offline), no auth, no real-time ingest, and the persona is desk-research rather than user-research. v2 starts with real analysts.

Product Decisions

Three trade-offs that mattered most.

01

Pre-select high-confidence detections as confirmed

Saves clicks on the easy 70% — but risks trust collapse if the model is wrong on something the UI presented as confident. Mitigated by logging every override.

02

Colour-code confidence; don't just show numbers

Pre-attentive processing is faster than reading under time pressure. Numbers stay available on hover for analysts who want them.

03

One-button dispatch, no confirmation modal

Every confirmed detection has already been reviewed individually. Adding a modal adds friction without adding information.

Restraint

What I deliberately didn't build.

Each of these would have been fast to build — and would have weakened the product:

  • A select-all button
  • Auto-dispatch above a confidence threshold
  • A confidence-tuning slider in the main UI
Measurement

The metrics I'd instrument.

Time-to-dispatch
North star — target <180s median
Analyst-confirmed precision
Counter-metric — target ≥95%
Override rate
On pre-selected detections — diagnoses model calibration
Flag rate
Calibrates how to surface uncertainty in v2
Next

What I'd build next.

  • Five real user interviews with analysts before any v2 spec
  • Override-driven model retraining loop
  • Multi-analyst handoff for flagged detections
  • Real-time ingest pipeline