Loading Smartgenhub
Products

Productised AI, powered by NVIDIA

Six packaged offerings built from systems we've already shipped and demoed. Each one runs on NVIDIA hardware — CUDA, TensorRT, Triton or Jetson — where our benchmarks justify it, and on plain CPUs where they don't. No shelf-ware: every product below is delivered as a scoped project with an evaluation report.

Document AI Vision Tagger Semantic Search AI Assistant GPU Serving Edge Vision The hardware behind them
Illustration of fields being extracted from a scanned invoice into structured data
Product 01 NLP · Document AI NVIDIA GPU fine-tuned

Document AI — Invoice & Form Extraction

Turns scanned invoices, forms and mixed-format PDFs into structured fields: vendor, dates, line items, totals. Built on a fine-tuned open model that flags uncertain fields for human review instead of guessing.

NVIDIA inside

Fine-tuned on a rented NVIDIA GPU; production inference runs on an L4-class card via TensorRT

You get

A deployed extraction API, review queue, and an evaluation report on your own documents

Measured accuracy

About 96% of fields read correctly on our 400-document test set — yours is measured before go-live

Known limit

Handwritten documents route straight to human review

Request a demo of Document AI →
Illustration of product photos being automatically tagged with categories and attributes
Product 02 Computer Vision CUDA · TensorRT

Vision Tagger — Product Photo Intelligence

Auto-tags catalogue and warehouse photos with categories, attributes and quality flags. Trained on NVIDIA cloud GPUs, served in batches on an inference card, and quantisable down to a Jetson module if tagging needs to happen on-site.

NVIDIA inside

Training on rented NVIDIA GPUs with CUDA; batch serving on an inference-class card, TensorRT-optimised

You get

A tagging API or batch job wired into your catalogue, plus per-class accuracy numbers

Deployment options

Your cloud account, our rented GPUs, or an NVIDIA Jetson box at the warehouse

Known limit

New product categories need a small labelled sample before they tag reliably

Request a demo of Vision Tagger →
Illustration of an AI workspace with a conversation panel and model selector
Product 04 LLM · Assistant GPU inference

AI Assistant — Private Workspace Copilot

A private assistant for your team: drafting, summarising, answering questions against your own data. Built on fine-tuned open models served from NVIDIA GPUs — in your cloud account or on your own hardware — so prompts and documents never leave your control. The same foundation as our own AI workspace.

NVIDIA inside

Open-model inference on NVIDIA GPUs; fine-tuning runs on rented cloud GPUs sized from benchmarks

You get

A deployed assistant with access controls, usage logging, and a written evaluation of its behaviour

Privacy

Self-hosted or your-cloud deployment; no third-party API is required for the core assistant

Known limit

It will make mistakes — we measure how often and design the workflow so they're cheap to catch

Request a demo of the AI Assistant →
Illustration of a containerised model API next to a monitoring dashboard with healthy status
Product 05 MLOps · Serving Triton · TensorRT

GPU Model Serving — Production Inference Stack

Takes any trained model — ours or one you already have — and ships it as a containerised API with logging, drift monitoring, cost tracking and rollback. NVIDIA Triton Inference Server handles GPU serving where the model earns it; the same container deploys to a CPU node when it doesn't.

NVIDIA inside

Triton Inference Server for GPU serving, TensorRT optimisation, and per-request GPU cost tracking

You get

A production API with monitoring dashboards, alerting, and documentation your engineers can maintain

Right-sized

Every GPU in the stack comes with the benchmark that justified it — no idle iron on your bill

Known limit

It's a hardened template adapted per project, not a one-click installer

Request a demo of GPU Model Serving →
Diagram of our hardware decision path routing workloads to CPU nodes, Jetson edge modules, or GPUs
Product 06 Edge · Vision NVIDIA Jetson

Edge Vision Kit — Inference Where Data Is Born

Vision inference next to the camera on NVIDIA Jetson modules (Orin-class): factory lines, kiosks, farms — places without reliable connectivity or where frames can't leave the site. Models are trained on cloud GPUs, then quantised and pruned to fit the module, with the accuracy trade-off measured before anything is committed.

NVIDIA inside

Jetson Orin-class modules running TensorRT-optimised models; training happens on rented NVIDIA cloud GPUs

You get

A configured edge box, an update pipeline, and the measured accuracy cost of quantisation

Best for

Sites with poor connectivity, strict data-residency needs, or per-frame latency requirements

Known limit

Module memory caps model size — some accuracy is traded for fit, and we show you exactly how much

Request a demo of the Edge Vision Kit →
Straight Answers

How buying one of these works

These aren't boxed software licences. Each product is a scoped delivery of a system we've built before, adapted to your data — priced after we've seen a sample, never blind.

1. See it run first

Every product demos live on a call, on real inputs. Bring a small, non-sensitive sample of your own data and we'll run it through while you watch.

2. Scoped, fixed deliverable

You get a written scope: what ships, what it's measured against, and what it costs — including the NVIDIA GPU hours, benchmarked and invoiced transparently.

3. You own the result

On final payment, the code, trained weights and documentation are yours. We're not an NVIDIA partner or reseller — we build on their hardware because our benchmarks keep pointing there.

Curious exactly which GPUs, CPUs and Jetson modules sit under these products? That's documented on our compute page.

Talk About a Product