PRD: Atlas — Analyst Triage Tool for AI-detected Objects

Author: Sai Gowrav P Status: Draft v0.1 — for prototype scope Last updated: [12-05-2026] Reading time: 8 minutes


TL;DR

Atlas is a web tool that lets an intelligence analyst review AI-generated object detections on overhead imagery, confirm or reject each detection in seconds, and dispatch a clean report downstream. It optimizes for time-to-dispatch under accuracy constraints — the analyst's bottleneck is not detecting objects (the model does that) but trusting and verifying them fast enough to matter operationally.

This document scopes a 2-week prototype, defines what's in and out, lists the trade-offs I made, and proposes the metrics that decide whether the next iteration is worth building.


1. Problem

1.1 The user

Intelligence analysts working with overhead imagery (satellite, drone) under time pressure. Their job is to identify objects of interest in incoming images and dispatch confirmed reports to operations teams within minutes of image arrival.

1.2 The current workflow (without AI)

  1. Image arrives in inbox / shared drive
  2. Analyst opens it in a desktop tool (often legacy)
  3. Manually scans for objects, draws bounding boxes
  4. Logs findings in a separate system
  5. Writes a report, sends it on

Pain: ~80% of an analyst's time is on the visual scan. Volume is rising; analyst headcount isn't. Latency to dispatch is operationally costly.

1.3 The opportunity

A reliable detection model can do step 3 in seconds. But analysts won't trust an AI-generated report they didn't verify. The product question isn't "can the model detect?" — that's an ML question. The product question is "how do we make verification fast enough that the AI is actually time-saving, not time-adding?"

1.4 Why this matters