Capabilities

Detection you can audit.
Evidence you can defend.

Vektra combines independent detectors, a verification layer, visual search and human review into one evidence chain. Every result keeps its source frame, the model that produced it, its scores and the reviewer’s decision.

Running in our recognition workspace

Eight capabilities.
One evidence chain.

From uploaded footage to a reviewed, audited record. Each stage keeps its own evidence, so every answer can be traced back to the frame it came from.

01 / MULTI-MODEL DETECTION

Independent detectors, compared openly.

Run detectors alone or in tandem on the same frames. Agreement is matched one-to-one by overlap and every model’s raw observations are kept and exportable, so agreement is a signal you can inspect rather than a verdict you have to trust.

  • Approved detector profiles, single or tandem
  • Strict one-to-one cross-model matching
  • Context classes recorded, never used to veto
02 / DETECT, THEN VERIFY

A second look at original resolution.

Candidates from the full frame are cropped from the original pixels and re-examined by a separately trained verification layer. Detector and verifier scores stay separate. Nothing is silently suppressed and there is no automatic all-clear.

  • Corroborated, uncertain and skipped outcomes
  • A verification budget per frame
  • A verification layer we trained ourselves
03 / VIDEO ANALYSIS

Every box tied to its moment.

Footage is decoded and sampled on its true presentation timestamps, and every detection is mapped back to native pixels. Detections of the same class are associated across frames, so persistence counts and a single frame never does.

  • Replay with overlays and a findings list that seeks to each moment
  • Temporal association that resets on scene change
  • Structured export of every observation
04 / VISUAL SEARCH

Find it again, and prove the match.

Footage and stills are indexed into private collections with self-supervised visual embeddings. A search region returns ranked candidates, each checked by keypoint and homography alignment. “Insufficient detail” is an acceptable answer, and a reviewer records accept, reject or unresolved.

  • Private, organization-scoped collections
  • Geometric verification on every candidate
  • Seek from a match to its source moment
05 / THREAT VOCABULARY

Built for protective security.

A security vocabulary covering firearms, edged and blunt weapons, explosives, and fire and smoke, with person and tool context to separate genuine threats from everyday look-alikes.

  • Multi-class threat detection
  • Context classes alongside threat classes
  • Profiles configured per deployment
06 / EVIDENCE INTEGRITY

Machine evidence that cannot drift.

File hashes are declared at upload and verified independently before analysis. Machine observations are immutable, human decisions are stored separately, storage is private with short-lived signed access, and every action lands in an append-only audit trail.

  • Hash verification at upload and at processing
  • Immutable observations and incident history
  • Append-only audit trail with trace identifiers
07 / MODEL INTEGRITY

Only approved models run.

Every model loads only after an exact cryptographic match against an approved manifest. There are no runtime downloads and no arbitrary uploads, and every detection records the model and hash that produced it.

  • Hash-pinned model manifests
  • Readiness reported from the model actually loaded
  • Model provenance on every result
08 / OPERATIONS

From a site plan to a resolved incident.

Sites, areas and camera inventory. Incidents that move from candidate to verified or dismissed, then acknowledged and resolved, with an immutable history. Role-based administration and secure, single-use intake links for onboarding new clients.

  • Multi-site inventory
  • Evidence-anchored incident workflow
  • Role-based access and audited sign-in
Human in the loop

The system proposes.
A person decides.

Vektra never turns a detection into an action on its own. Candidates, verifications and visual matches arrive as separate, labelled signals. A reviewer confirms, dismisses or leaves the question open, and that decision is recorded alongside the machine evidence, never on top of it.

  • 01Machine evidence and human decisions stored separately
  • 02No fused score and no automatic all-clear
  • 03Uncertainty and “insufficient detail” reported, not hidden
  • 04Alerting enabled only by explicit, supervised configuration
One incident, end to end

Follow a single clip
from upload to record.

Every stage writes its own evidence. Watch the record on the left fill in as one incident moves through the chain.

How a model layer is built
01 / Upload

A clip arrives.

Its hash is declared at upload and verified independently before analysis. Frames are sampled on their true timestamps.

02 / Detect

Two detectors look at the same frame.

They run independently. Their raw observations are kept, matched one-to-one by overlap where they agree, and never used to veto each other.

03 / Verify

The candidate gets a second look.

It is cropped from the original pixels and re-examined by a separately trained verification layer. Outcomes are corroborated, uncertain or skipped; nothing is silently suppressed.

04 / Associate

The same object, frame after frame.

Detections of the same class are associated across frames into a track that resets on scene change, so persistence counts and a single frame never does.

05 / Search

Has this been seen before?

A region is searched against private collections with visual embeddings; each candidate is checked by keypoint and homography alignment. Insufficient detail is an acceptable answer.

06 / Review

A person decides.

Candidates, verifications and matches arrive as separate, labelled signals. A reviewer confirms, dismisses or leaves the question open, and that decision is stored apart from the machine evidence.

07 / Record

The record holds.

Observations are immutable, the incident history is append-only, and every entry carries the model and hash that produced it.

Where it runs

Cloud GPU today.
Edge by design.

The workspace runs as containerized GPU services with durable, lease-fenced job queues and private storage. The same detectors and evidence model compile to TensorRT engines for Jetson-class edge modules on aircraft, vehicles and site gateways.

  • 01Containerized inference, video and recognition workers
  • 02Readiness attested from the loaded model
  • 03Organization-scoped data with server-side access only
  • 04TensorRT compilation for edge targets
Built around your challenge

What do you need
to recognize?

Start a project