See who sees you. Automated surveillance detection for cities, routes, and protectees.

Intelligence officers, diplomats, judges and witnesses move through cities saturated with cameras. Hostile services also move through those cities — in cars, in pairs, in floating boxes — following the people they want to compromise, coerce, or harm. The tradecraft for detecting a tail is called a Surveillance Detection Route, and it has existed for 80 years. It is manual, expensive, and does not scale. Backtrack automates it. Public and ministerial cameras along a pre-registered SDR corridor are fused into one co-travel engine. Every vehicle that hits enough of your chokepoints, too close to your run, is scored. Every score comes with the evidence trail. Your officer gets the call before the adversary does.

Surveillance is a pipeline. Every stage leaks.

A hostile intelligence or organized-crime surveillance operation is not a single car in the rear-view mirror. It is a pipeline of teams, vehicles and observation posts — trained to swap positions, change plates, fall back, leapfrog parallel streets, and pre-position at your likely destinations. Ukraine 2014–2025 and the Mexican cartels' counter-surveillance doctrine are the two richest contemporary case-libraries; the tradecraft itself dates to Cold War Berlin and Moscow. Every stage of the pipeline leaves a trace on the city's camera fabric. Without an engine to fuse it, those traces die in siloed DVRs. Backtrack is that engine.

Every service protecting field officers, every close-protection detail for dignitaries, judges and witnesses, and every counter-terrorism unit tracking pre-attack surveillance faces the same structural problem: the signal is in the city's cameras, but nobody is fusing it in time to matter. Backtrack fuses it.

Backtrack is not an ALPR vendor. It is tradecraft in a binary.

Four tradecraft principles guide every design decision. An ALPR vendor sells you plate reads. A correlation vendor sells you dashboards. Backtrack sells you an automated surveillance-detection officer — something that understands what an SDR is for, why static OPs matter, why one plate on three unlikely cameras is signal and ten plates on one boulevard is noise. The craft is older than the code.

One map. Every camera. Every plate. Four detectors. One score.

At the unit level: a correlation service plus a fleet of edge ALPR + re-ID nodes installed at ministerial, municipal and transit camera feeds, plus a hardened operator app on a vehicle tablet and a command console for the surveillance-detection coordinator. At the system level: a real-time graph over the city's camera fabric that answers one question — who is co-travelling with the protectee right now, and how unlikely is that? — and delivers the answer, with evidence, before the run is over.

         ┌───────────────────────────────────────────────────┐
         │  BACKTRACK · SDR RUN ECHO-9 · T+18:42              │
         │                                                   │
         │   ┌──────────┐     ◎ Protectee  CA 8421 AK         │
         │   │  Edge    │──── plate: CA 3847 PR              │
         │   │  ALPR    │     reid:  black SUV · roof rack    │
         │   │  N=120   │     score: 4/5 chokepoints          │
         │   └────┬─────┘                                     │
         │        │       ┌───────────┐                       │
         │   ┌────▼────┐  │ Correlate │                       │
         │   │  Graph  │──│  P = 2e-6 │  → SCORE 91        │
         │   │  engine │  │  α = 0.01 │                       │
         │   └────┬────┘  └───────────┘                       │
         │        │                                            │
         │   ┌────▼────┐         ┌──────────┐                │
         │   │  Evid   │ ──────► │ Operator │  TABLET          │
         │   │  trail  │         │  tablet  │                │
         │   └─────────┘         └──────────┘                │
         │                                                   │
         │  transport: Nexus Atlas · bonded LTE/mesh/SAT        │
         └───────────────────────────────────────────────────┘
      

Five phases. Minutes, not days.

A Backtrack run is a pre-authorized, time-bounded, scope-limited query against the city camera graph. Authorization is explicit (the protectee's controlling service), scope is narrow (only the run's corridor and window), purpose is declared (counter-surveillance for the named protectee), and every query is logged immutably for the inspector general. The system is useless without this envelope — legally, operationally, and ethically.

Four tiers. Every one runs on-prem or on bonded transport.

Backtrack runs on Nexus Atlas as its transport fabric. That matters: the operator tablet must work in a parking garage, on an underpass, under cover of a dense urban RF environment, or when the adversary is actively jamming LTE. Bonded LTE + mesh + SAT means the alert arrives on whichever path survives. No cloud dependency, no vendor phone-home, no plate read ever leaves the customer's legal jurisdiction unless the customer explicitly pushes it.

  Tier 01 · Cameras                Tier 02 · Edge                Tier 03 · Correlation            Tier 04 · Operator
  ┌──────────────────┐             ┌──────────────────┐           ┌──────────────────────┐       ┌──────────────────┐
  │ MoI PTZ          │  RTSP       │ Edge node A      │  plate +  │ Graph engine          │  NXA  │ Field tablet     │
  │ transit ANPR     │ ──────────► │ 8 feeds · NPU    │ ────────► │ 4 detectors, hot/hot  │ ────► │ alert · map       │
  │ municipal fixed  │             │ ALPR · re-ID · P │  embed    │ PoL baselines         │       │ loss drills      │
  └──────────────────┘             └──────────────────┘           └──────────────────────┘       └──────────────────┘
           ▲                                 ▲                              │                              ▲
           │                                 │                              │                              │
           └── camera feed stays on          └── only hashes & embeddings   └── signed evidence ──────────┘
               camera's own retention            leave the edge                 packet + IG audit stream
  

One score. Four orthogonal evidence sources.

A professional surveillance package defeats any single detector on its own. Co-travel misses the floating box. Floating-box misses the static OP. Static-OP misses the plate-swapped clean vehicle. Four detectors run in parallel and their scores fuse by calibrated p-value combination. A hit on two is evidence; a hit on three is a call. Every point of every score traces to a camera-timestamp pair — the score is not a black box.

Score fusion

Each detector emits a calibrated p-value under its null. Fused using Fisher's combined probability test, with correction for non-independence when detectors share evidence (co-travel and set-level share chokepoint hits). Threshold defaults to p < 10⁻⁴; tunable per mission criticality. The score shown to the operator is 0–100 mapped from log-p, for readability, but the underlying decision is always on the calibrated p-value.

Plates lie. Vehicles don't — not as easily.

A plate is a string. A vehicle is a fingerprint: make, model, color, body-panel signature, wheels, roof rack, antenna mount, sunroof, dent pattern, aftermarket bumper. The same vehicle moving through different cameras produces similar embeddings from a re-ID model even when the plate has changed. Backtrack co-indexes plate reads with re-ID embeddings. A plate with no matching visual, or a visual with no matching plate, is itself a red flag — and it is exactly what a professional plate swap looks like.

08:42  NODE-12  plate KN 9102 VT  visual: silver VW Passat
09:14  NODE-18  plate KN 9102 VT  visual: silver VW Passat
11:02  NODE-03  plate KN 9102 VT  visual: black SUV   ← mismatch
11:47  NODE-07  plate KN 9102 VT  visual: black SUV
13:18  NODE-14  plate CA 3847 PR  visual: black SUV   ← swap complete
14:02  NODE-22  plate CA 3847 PR  visual: black SUV
        

What the fingerprint actually contains

Vehicle re-ID is not a single embedding — it is a composite of seven attribute families, each independently scored and independently checkable by an analyst. "Same vehicle" is an argument with thirty lines of evidence, not a single cosine distance. Each attribute contributes 1–10 bits of identifying information; across 20+ attributes you reach 40–60 bits — enough to pick out one vehicle from a city of 500K. That is the evidential math behind every re-ID call.

Analyst-only cues · surveillance-vehicle tells

Some observable signals are too weak or too context-dependent to auto-score, but they matter to an experienced analyst. Backtrack surfaces them as evidence rows in the review queue — never as live alerts — so the human call is informed without the automation over-reaching. Operators train against this list; the system records it.

Normality has a shape. Surveillance breaks that shape.

For every camera on the network, Backtrack maintains a rolling 60-day baseline of who appears there, at what hours, on what days, with what vehicle class. The baseline is a probability distribution, not a whitelist — no vehicle is "known-good," only known-typical. When a specific plate or visual fingerprint appears at a camera at a time it almost never appears, the deviation is measured in standard deviations and scored against the protectee's movement. Pre-positioning surveillance, which ignores the target's movement and waits for it, lights up only on this detector.

The officer is driving. The alert is a sentence.

An operator tablet that demands reading while the officer is negotiating a turn is a liability, not a tool. Backtrack's alert path is voice-first (ear-piece TTS with a fixed phraseology), tablet-second (map overlay + evidence card), and console-third (full graph at the SD coordinator's desk). The officer gets a sentence. The coordinator gets the forensics. Both surface the same evidence packet so no two people in the loop see a different picture.

Your own PPD looks exactly like a tail.

Most of the vehicles that will appear to be tailing a protectee, in any realistic deployment, are friendly — the same ministry's protective detail, allied-service liaison, sometimes another department's surveillance of an entirely different target who happens to share a corridor. A Backtrack that flags them all is worse than useless: it will be ignored within a week, and it will leak operational information about who-is-protected-when to anyone inside who queries the log. Deconfliction is the core operational problem, not an afterthought.

This is how the system stays legal. And how we stay the good guys.

A real-time fusion engine over a city's cameras is, in the wrong configuration, exactly the mass-surveillance infrastructure we refuse to build. The differentiators that keep Backtrack a legitimate counter-surveillance tool rather than a regime-accessory are not optional features — they are how it's architected at the schema level. If a prospective customer asks to remove any of these, we decline the pilot. We have.

Foot-mobile surveillance detection. Walled off for a reason.

Hostile services run foot tails constantly — in malls, transit stations, squares, narrow old-town streets where vehicles cannot follow. A case officer who drives a clean vehicle SDR can still pick up a three-person rotating foot team the moment they step out of the car. Backtrack-Foot is the pedestrian analogue of the vehicle core: the same four detectors (co-travel, set-level, parallel-route, static-OP) applied to whole-body appearance embeddings instead of plate reads.

It is not part of the default deployment. It ships as a separately licensed, separately authorized module under stricter constraints than the vehicle core, because pedestrian appearance embeddings are biometric data under GDPR Article 9 and LED Article 10. The reason this section exists at all is that foot-mobile surveillance is too operationally important to ignore. The reason it is walled off from the core is that "whole-body biometrics" is the line that separates a CI counter-surveillance tool from a mass-surveillance suite. We do not cross that line by accident.

What the foot fingerprint contains

Deliberately narrower than vehicle re-ID. No face. No gait. No iris. No ear shape. No voiceprint. We take the appearance cues that survive short-term clothing variation — because those are what matter for catching a tail during a run — and we stop there.

Analyst-only cues · probable foot tail

Like the vehicle analyst-only list, these are cues too weak or too context-dependent to auto-score but valuable to an experienced SD analyst. Surfaced as evidence rows, never as officer alerts.

Foot-specific guardrails · in addition to core guardrails (§11)

Built for the field officer. Scales to the protectee.

The primary deployment is counter-intelligence and counter-surveillance for ministry-of-interior officers operating in-country. The secondary applications, which share the same engine with scoped authorization, are close-protection for dignitaries, judges and witnesses — plus counter-terrorism pre-attack surveillance-detection. The analytics are identical; the authorization, the allow-list and the DPIA differ.

Primary · intelligence officer protection

Secondary · dignitary, judicial & witness protection

Secondary · counter-terrorism pre-attack indicators

What "good" looks like, when it works.

These are design targets for the correlation node, the edge node and the operator app. All subject to validation in a pilot deployment. Nothing on this table has been measured on real field hardware. We publish the targets so they can be argued with, not because they are a promise.

There are five adjacent categories. We are none of them.

Three curves cross in 2026.

Pilot, harden, scale.

Who we want to talk to.

Backtrack is pre-prototype. Before we write code for the correlation engine we want to settle the DPIA template, the deconfliction protocol, and the operator phraseology with real field users. If you run a CI directorate, a PPD, a witness-protection unit, a CT fusion cell, or a judicial-protection office, and the problem on this page matches a problem on your desk, we would like to talk.