● 100% Non-Red Supply Chain · Made in Taiwan · CTE Group

Drone vision that sees in the dark — the AI ISP, reinvented in software.

AI Angel builds a non-red AI image-signal-processing (ISP) and edge-AI platform for drones — edge denoising and edge-AI scene understanding today, with native wide dynamic range (DOL HDR) on the roadmap, and verified over-the-air model updates. Proven on FPGA today, advancing to a dedicated 28 nm AI ISP chip. With partner JANMAN Precision, the rugged airframe and the AI brain arrive as one turnkey package.

Talk to us Explore the technology

120 dB class HDR — DOL HDR, roadmap

Field-upgradable edge AI · OTA

Data stays on-device — offline inference by design

Dual integrity check — publisher hash + on-device checksum

One product, two disciplines

The AI brain and the rugged airframe — from one partnership.

AI Angel (CTE Group) provides the intelligence; JANMAN Precision provides the structure. Together, one complete, fully non-Chinese package.

The intelligence — AI Angel / CTE Group

Sees in the dark

AI ISP: edge denoise and dehaze with AI scene recognition; DOL HDR on the roadmap
Verified OTA — dual integrity check; failed updates are rejected and rolled back automatically
Offline inference — raw imagery never leaves the aircraft
Built non-red: Lattice FPGA today → dedicated 28 nm AI ISP chip

The structure — JANMAN Precision (建名精密)

Survives the field

Aerospace-grade PEEK / CFRTP rugged housings; anti-vibration, IP-rated
One-stop: mould → injection → metal insert → assembly
Precision to 0.0005 mm · 56 patents · founded 1985
ISO 9001 / 14001 / 13485 · MoHW GMP · LEED green factory

Why one partner

One stop — the airframe and the AI, delivered as one.

Software and hardware from two different industries, yes — but you don't integrate them yourself. You buy a pre-matched unit from a single Taiwanese partnership: one purchase, one integration, one point of accountability.

🛒

One purchase, one PO

The rugged airframe, the edge-AI brain and the OTA service arrive as a single package — no juggling a separate hardware supplier and a software supplier, or making them agree.

🧩

Pre-integrated & co-designed

Enclosure, sensor mount, FPGA and AI models are matched and validated together. You receive a working unit — not an integration project to run yourself.

🤝

One point of accountability

A single partnership owns the whole stack — one warranty, one SLA, one contact. If anything needs attention, there's no finger-pointing between a hardware and a software vendor.

🛡️

One clean non-red supply chain

Hardware and silicon are both non-Chinese, sourced through one trusted partnership — simpler compliance and procurement for government and critical-infrastructure buyers.

⚡

Faster to the field, lower risk. Skip the multi-vendor integration cycle — a pre-matched hardware + AI unit gets you to a working deployment sooner, and the same partner keeps the AI improving over the air across the platform's life. Hardware and intelligence stay in step, from one source.

The technology

A full AI ISP engine — algorithm and silicon, co-designed.

Eight tightly-integrated capability areas turn a raw, noisy sensor feed into clean, decision-ready imagery — efficient enough to run on the edge.

🧩

Architecture & AI accelerator

ISP pipeline and memory subsystem in FPGA, co-designed with a separate edge AI accelerator across a standard interface; deterministic data paths, configurable pipeline.

🌅

ISP core algorithms

Spatial/temporal denoise, dehaze, local tone mapping and colour processing shipping today; multi-exposure HDR and motion compensation on the roadmap.

🧠

AI models & optimisation

Lightweight CNNs, INT8 quantisation, knowledge distillation and hardware-aware training — big-model accuracy at edge cost.

⚙️

Hardware acceleration

Dedicated ISP accelerators — feature detection, motion-estimation and filter arrays, colour-transform engines.

🔋

Memory & low power

Hierarchical cache, DMA and bandwidth optimisation; DVFS, clock-gating and power-domain design targeting < 2.5 W on 28 nm (roadmap · pre-silicon estimate).

🛡️

Update integrity & model protection

Every update package is verified twice — publisher hash plus on-device checksum; any mismatch is rejected and rolled back automatically. Model IP is protected by a three-layer scheme: contract terms, per-customer watermarking and model freshness — the continuously-improving loop devalues any static copy. Inference runs offline by design.

🧩

Software-defined hardware — not frozen silicon

Our vision engine runs on FPGA, so the logic itself is reconfigurable. Where other vendors hard-wire their circuits into a fixed chipset (ASIC) and can never change them in the field, we update both the AI models and the hardware logic over the air — the platform keeps evolving long after purchase.

System architecture

One integrated stack — from the sensor to the cloud.

Image quality, edge AI and secure updates are co-designed as a single system. The drone stays fully autonomous in the field; the cloud keeps the whole fleet improving over the air.

100% Non-Red Lattice FPGA AI inference · dedicated accelerator Offline inference · data stays local OTA hot-swap · no downtime On-chain provenance

① Cloud & OTA platform

Autonomous agent

🤖 Self-managing OTA

An AI agent continuously watches live quality metrics — PSNR, SSIM, latency, accuracy — and decides on its own when to push, upgrade or roll back a model. No truck-roll, no downtime.

Platform

☁️ Model & fleet management

Version control, A/B deployment, knowledge-distillation training and a live dashboard across the entire fleet.

Data & trust

🔐 Data residency & audit

Raw imagery never leaves a device. Federated learning — sharing only de-identified gradients — and a tamper-evident ledger for deployment audit are roadmap items.

Encrypted transport between cloud and fleet — OTA channel over 4G (AES-128 transport encryption), zero raw imagery in transit ▼

② Edge device · on the drone

Image input

📷 Sensors → FPGA

Multi-camera RAW feed streamed straight into the FPGA over MIPI CSI-2 — no host PC in the loop.

Image engine · core IP

🧩 Lattice FPGA image engine

A full 9-stage ISP pipeline runs entirely in FPGA fabric with deterministic timing — ISP-block latency on the order of 15 µs — and boots without a host. Measured drop-free at 1920×1080 @ 60 fps. End-to-end system latency is counted in frames and is defined jointly during technical integration. The AI perception layer runs on a dedicated accelerator, decoupled from the FPGA bitstream and updatable over the air — the proprietary ISP core that anchors our roadmap to a dedicated AI ISP chip.

9-stage ISP pipeline

1Black level

2Digital gain

3Defect fix

4Demosaic

5HDR

6Colour matrix

7Colour space

8Gamma

9AI hand-off

Stage 5 (HDR) is present in the pipeline but bypassed in the current single-exposure build; DOL HDR is a roadmap item.

3-tier AI perception layer · on a dedicated accelerator

Tier 1 · Scene controlAdapts the ISP to the scene — models updatable over the air.

Tier 2 · EnhanceDenoise today; super-resolution and HDR fusion on the roadmap — all decoupled from the FPGA bitstream.

Tier 3 · PerceptionObject detection + depth, run asynchronously on the accelerator.

Dedicated flight controller · verified OTA

⚙️ On-device flight brain

A dedicated flight controller drives the FPGA and manages A/B firmware: it evaluates each new model against the quality metrics, switches atomically with image-output freeze far below the acceptance limit, and rolls back automatically if anything regresses — the drone keeps flying through every update.

FPGA logic updates · trial, then commit

🔁 Bitstream update safety

FPGA bitstream updates are two-phase — trial, then commit. A trial image is verified in a separate slot; power loss during trial automatically reverts to the current version. If a commit write is interrupted, the device boots the write-protected golden image. Built on the FPGA vendor’s native configuration architecture — an update can never leave the vehicle without vision.

Three independent radio links

2.4 GHz · ELRS

Flight control (C2)Dedicated command & control link, 10 km+ range.

4G LTE

Backup + OTAModel updates, telemetry and image upload — kept off the control link.

5.8 GHz

FPV videoPilot's live feed on its own link, so nothing competes for bandwidth.

Performance

Proven on FPGA today.

Demonstrated results, measured on the FPGA platform — the same figures we accept a system against.

Every performance claim on this site is labelled at one of three levels: Measured (with a measurement record), Derived (with a stated calculation basis), or Roadmap (design frozen, not yet built). At acceptance, every number can be re-measured on site by your team.

Proven today

Validated results

ISP pipeline · on FPGA9-stage · 1080p 60 fps, drop-free

Cold-start self-load10 / 10 validated

Denoise · offline validation−45% / −49%

Distillation (teacher / student)34.42 / 33.89 dB

OTA model updatehot-swap + rollback

Raw data leaving devicenone

ISP block latency15 µs class · deterministic

Model OTA loop · on hardwaredeploy → verify → A/B switch → rollback

A/B switch freezefar below the 50 ms acceptance limit

Edge denoise gain · on-device+6.4 dB

Model size · 4G deliveryKB-scale · < 1 s

The model OTA loop is verified on hardware: remote deploy → dual integrity check → atomic A/B switch (freeze time far below the 50 ms acceptance limit) → automatic rollback on failure. Edge denoise gain +6.4 dB measured on-device; models are KB-scale, delivered over 4G in under a second.

Measured: ISP pipeline, cold-start, ISP block latency, OTA loop, edge denoise gain, model size and delivery time. Derived (offline, 600 frames): −45% / −49% denoise and the 34.42 / 33.89 dB distillation figures. Roadmap: 120 dB DOL HDR and the < 2.5 W 28 nm power target.

Proven capability

AI Image Denoising — Running on the Edge NPU

Same source footage, side by side. Our distilled denoising model reduces spatial noise by 45% and temporal noise by 49%, with no loss of edge detail.

The model is KB-scale in INT8 — small enough to run on the attached edge NPU alongside the FPGA ISP pipeline, delivered over 4G in under a second and swappable over the air.

−45%Spatial noise

−49%Temporal noise

KB-scaleINT8 model

Denoise figures are offline-validated on identical material (600 frames); on-device measurement on the edge NPU is reported in the Performance section.

ISP · 9-stage pipeline, 1080p 60 fps drop-free

Boot · FPGA cold-start self-load from SPI flash · 10/10 validated

Platform · Lattice ECP5-85F

Ownership model

Buy the hardware once. The AI keeps getting better — over the air.

One rugged platform, a decade of upgrades. You pay for the hardware once; the intelligence keeps evolving through verified OTA — no forced refresh, no downtime.

DAY ONE ONE-TIME

Hardware — a single purchase

Rugged JANMAN airframe + FPGA edge-AI. One capital investment, yours to keep — no per-seat or per-frame licence.

CONTINUOUS

Unlimited AI updates, OTA

New denoising and perception models pushed over the air — integrity-verified, atomic A/B switch, automatic rollback on failure — as often as we improve them.

EVERY RELEASE

The fleet keeps improving

The same drones get sharper each update — better low-light, new capabilities — without ever touching the hardware.

10-YEAR LIFECYCLE

One platform, a decade of AI

A full decade of upgrades on one platform — no hardware refresh cycle. The drone you buy today keeps getting smarter.

💡

The pricing model: one hardware purchase, unlimited AI updates.

A single, one-time hardware investment. From day one, the AI keeps improving over the air — unlimited updates across the platform's 10-year life, with no forced hardware refresh. Buy the metal once; the intelligence never stops evolving.

Data sovereignty & OTA

Your data never leaves. The AI still gets better.

“Data Sovereignty.”

The most common question about over-the-air AI is whether it means your imagery leaves the site. It doesn't. "Over the air" is about updates arriving — not footage departing. Both your raw imagery and your training data stay inside your environment.

Inbound

Updates come to the device

An OTA update is model weights pushed to the aircraft — integrity-verified on arrival, atomic A/B switch, automatic rollback. No customer data has to leave for an update to happen.

Outbound

Your training data stays internal

Training runs inside your environment (on-prem), so your training data is never uploaded. Federated fleet learning — sharing only de-identified gradients, encrypted math and never pictures, and only with your consent — is a roadmap item.

Zero-egress option

Fully sovereign, nothing leaves

For the highest-security sites, a fully offline deployment sends nothing — not even gradients. The model is improved off-site and delivered to you as scheduled updates, each verified on-device before it takes effect, and logged.

🛡️

"Your images — and your training data — never leave your site. Only the model's learning, as encrypted math (never pictures), is shared, and only if you allow it. For fully sovereign deployments, we bring you scheduled updates with zero data egress."

Manufacturing partner

JANMAN Precision — aerospace-grade hardware, made in Taiwan.

The rugged airframe, enclosure and precision mechanics that keep the AI stable in the field — delivered as one-stop, mould-to-assembly production from a vertically integrated Taiwanese manufacturer.

🧬

Super-engineering plastics

Medical-grade PEEK / PPSU / PEI moulding — high-temperature, radiation-resistant and low-outgassing. "Plastic-for-metal" lightweighting that keeps airframe rigidity while cutting weight.

📐

Micro-precision machining

Sodick wire-EDM and Makino high-speed milling to 0.0005 mm precision — for gimbals, camera mounts and precision actuator mechanisms.

📡

RF antenna & connector processes

Ultra-fine wire over-moulding (Ø < 0.35 mm) and metal insert-moulding — the core craft behind high-frequency video links, receivers and array-antenna connectors.

🧪

Cleanroom & scientific QC

ISO Class 7 / 8 cleanroom moulding; ZEISS CMM and NIKON vision metrology delivering full GD&T data-driven inspection reports.

🏭

One-stop production

Mould design → mould-flow analysis → injection → metal insert → cleanroom sub-assembly and ESD-protected testing, under one roof.

🛡️

Proven quality system

ISO 9001 / 14001 / 13485 and MoHW GMP; recognised with the Yushan National Brand Award and the Dun & Bradstreet SME Elite Award.

Inside the factory

A vertically integrated plant — from mould shop and machining to injection moulding, cleanroom assembly and metrology.

Wire-EDM & precision machining shop — Wire-EDM & precision machining

1985
Founded

NT$200 M
Capital

136
Specialists

56
Patents

LEED
Green factory

ISO 13485 · GMP
Certified

The company

Two decades of imaging silicon — now aimed at drones.

AI Angel · CTE Group

Part of CTE Group (founded 2010, Taipei) with deep heritage in image-signal-processing and camera-module design across mobile, wearable, automotive and security — and ASIC & FPGA imaging experience. AI Angel focuses that expertise on non-red drone edge-AI, OTA and a dedicated AI ISP chip, delivered ODM/OEM with an open SDK to empower local manufacturing.

JANMAN Precision · 建名精密

An aerospace-grade precision manufacturer founded in 1985, with NT$200 M capital, ~136 specialists and a LEED-certified, vertically integrated factory. Core strengths: super-engineering-plastic (PEEK / PPSU / PEI) moulding, micron-level machining and one-stop mould-to-assembly production — backed by 56 patents, ISO 9001 / 14001 / 13485 & MoHW GMP, and the Yushan National Brand Award. JANMAN provides the rugged structure that keeps the optics stable in the field.

Get in touch

Let's talk about your drone programme.

Tell us your use case and we'll come back with a concrete next step — a demo, a Sample-Kit pilot, or an ODM/OEM discussion.

Benison.pin@aiangel-ctegroup.com.tw
benison.pin@gmail.com

Company

AI Angel · CTE Group (晶芯智創), Taiwan
in partnership with JANMAN Precision (建名精密)

Contact

Benison Pin (賓振呈) — CTO

Website