The Thesis

Fig. 01

OUTPOST — MAHAMAIA

The uplink is scarce.
The intelligence doesn't have to be.

Video doesn't fit over Starlink, LTE, or VSAT at scale. OUTPOST is an edge AI transport gateway: it converts camera feeds into bandwidth-efficient, detector-compatible streams, so your AI can operate continuously — not only when an event is detected. A stream sized for what your models actually consume.

Fig. 01 — shipping today: OUTPOST Bridge

The Measurement

Continuous AI ingestion, measured on real footage.

51 maritime videos, 11,860 detection boxes, split-by-video to prevent data leakage. OUTPOST ships today as a transport gateway. The next generation replaces video transport with learned representations.

Shipping today — OUTPOST Bridge

Continuous AI ingestion over a constrained link. Measured real bytes, no detector changes, no new hardware.

Operating point	Bitrate/camera	Detection recall	Sub-16px recall
256×256, CRF 32, 10 fps	54 Kbps	86.4%	71.7%

16 cameras ≈ 0.86 Mbps aggregate at the same operating point (extrapolated from the per-camera measurement above). Designed for continuous detector-resolution ingestion over constrained links.

In development — OUTPOST Representation Transport

The goal: don't send pixels across the link. Send the features the models need instead. Target: ≥2.6× smaller than H.265 at equivalent detection, evaluated against H.265 QP22. Benchmarks pending real-byte validation.

Method	Bitrate	Detection recall	Sub-16px recall
OUTPOST Representation Transport (target)	≤ 220 Kbps (CDF)	—	≥ 87.4%
H.265 QP22 (measured baseline)	564 Kbps	87.8%	—

Bitrate figures for the shipping bridge are measured real bytes; development-target figures are CDF estimates until production coding is validated.

III

The Platform

Architecture

A codec is one component. OUTPOST is the system.

Today, every model decodes the same video and processes it again. The transport gateway solves the bandwidth problem today. The representation layer is how we eventually remove video decoding from the path.

Shipping today

Edge AI transport gateway — RTSP in → detector-compatible H.265 → constrained link → decode → existing AI system. One container per gateway.

Roadmap

Representation encoder → packet format → publish/subscribe broker → SDK. “New models ship as subscribers. The edge never changes.”

Fig. 02 — next generation: representation transport (roadmap)

	Traditional codec	OUTPOST — next generation
What travels over the link	Compressed video	Machine representations, not pixels
Decode at the cloud	Every consumer decodes independently	No decode — representations become direct model input
Add a new AI model	Provision new decode + backbone pipeline	Subscribe to the existing stream
Bandwidth under load	Scales with resolution × frame rate	Scales with the information required by downstream models
Edge hardware change	Rebuild and redeploy the pipeline	Encoder upgrades transparent to subscribers

Core components

01
Transport Gateway
— Shipping
RTSP in → detector-compatible H.265 → constrained link → decode → your existing AI system. One container per gateway.
02
Representation Encoder
— Roadmap
Runs at the edge. Converts frames to feature packets once.
03
Packet Format
— Roadmap
Versioned wire format. Immutable — anything built on it keeps working.
04
Broker
— Roadmap
Publish/subscribe transport. Fan-out, backpressure, access control.
05
SDK
— Roadmap
Python and C++ APIs. Task heads plug into the same abstractions.

New models ship as subscribers. The edge never changes.
— Design principle, next generation

The Pilot

Fig. 02

Prove it on your link, on your footage.

A structured pilot on your deployment: we instrument your cameras and your uplink, then benchmark OUTPOST against your current codec under identical conditions. You receive the complete detection-vs-bitrate curve for your own footage — not ours.

The result answers the question that matters operationally: how many cameras can your existing link carry at usable detection range?