Cookies on this site

We would like to use cookies to understand how the site is used and to measure our campaigns. Nothing is set until you choose, and you can change your mind at any time from “Cookie settings” at the foot of every page. Privacy policy

Contact UsContact Us

Services / AI Consulting / Hybrid & Edge AI

Run AI where the data lives — on the device, on your servers, or in the cloud.

Not every workload belongs in a cloud model. We design routing policies that keep sensitive work local, send only what needs scale to the cloud, and prove the cost and latency difference before you commit.

Book a hybrid AI assessmentRead the bid-triage case study
11×
faster: 30B local model vs 120B, identical verdicts
0 bytes
of document text sent to a cloud API in our bid-triage pipeline
3 tiers
device · your infrastructure · cloud, chosen per task

When it fits

Three reasons to keep AI close to the data

If any of these describe your workload, a hybrid design will usually beat a cloud-only one.

  • Regulated or sensitive data

    Certificates of insurance, patient records, contracts, source code. Redact or extract on-device and send only sanitized output onward.

  • Latency and offline use

    Field teams, branch offices, factory floors and kiosks that cannot wait on a round-trip or a connection.

  • Cost at volume

    Thousands of classifications a day add up. A small model on a $3k box beats per-token pricing — if it is accurate enough. We measure that first.

How it works

One policy decides where each task runs

Every request is scored on sensitivity, latency, cost and required capability, then routed. Fallbacks are explicit, and every hop is logged.

  1. 01 On-device

    Laptops, mini PCs, edge boxes

    Classification · redaction · extraction · transcription · embeddings · intent

    Foundry Local · Windows ML · Ollama · NVIDIA GB10

  2. 02 Your infrastructure

    GPU servers, Azure Local, Kubernetes

    RAG over governed content · batch extraction · model serving near regulated systems

    Azure Local + Arc · Foundry Local · vLLM

  3. 03 Cloud

    Azure AI Foundry, Azure OpenAI

    Large-context reasoning · multimodal · orchestration · agent fleets · monitoring

    Azure OpenAI · Copilot Studio · Agent 365

Routing policy

  • Sensitivity → stays local
  • Latency budget → nearest tier
  • Cost ceiling → smallest model that passes
  • Capability → escalate with context stripped

NVIDIA DGX Spark

A desktop AI system for departmental document work

NVIDIA DGX Spark and the partner desktops built on the same GB10 chip run models large enough for real document work, on a desk next to the people who use them.

Our own bid desk is the proof. Every morning a pipeline reads about 2,000 new listings from government procurement portals, parses the bid documents and emails a short list of the 20 to 50 worth quoting. Every model call runs under Ollama on one NVIDIA GB10 desktop system in our office, and no document text is sent to a cloud model.

The model is a 30B Nemotron. On the same bids, with byte-identical inputs, it was 11 times faster than a 120B model and reached the same verdicts, so the larger model stays unused. We size a desktop system for a client the same way: benchmark the task on your real documents, pick the smallest model that passes, and keep the cloud for the steps that need its scale.

How the bid-triage pipeline works
  • Batch document work for one team: triage, extraction and classification
  • A 30B model in daily use on our own GB10 desktop
  • Sensitive documents stay on hardware you control
  • Measured against a cloud baseline before you commit
  • Hardware sourced, set up and supported by our group

Proof

What we have shipped on local models

Internal builds and client demos, measured against cloud baselines.

All case studies
  • Internal project

    2,000 → 50 a day

    Public-sector sales

    Hybrid & Edge AI

    Procurement triage on a local 30B model

    Logs in to five government procurement portals each morning, reads about 2,000 listings, parses the bid documents and emails a short list of the 20 to 50 worth quoting. Every model call runs on one desktop AI system in our office.

    • 2,000 listings down to about 50 a day
    • 11× faster than a 120B model, same verdicts
  • Product

    Every prompt screened

    Any industry on Microsoft 365

    AI Safety

    AI activity and data posture for Microsoft 365

    Sentinel watches Copilot and Copilot Studio agents in a Microsoft 365 tenant. It logs prompts and tool calls, scores each agent's risk, flags oversharing and stale access, and screens prompts for injection attempts with a classifier that runs locally at no per-call cost.

    • Every agent prompt screened
    • Runs inside the customer's own Azure tenant

Questions

Common questions

What CTOs and IT directors ask us first.

Stack we work with

  • Foundry Local
  • Windows ML
  • Azure Local
  • Azure AI Foundry
  • Copilot Studio
  • Ollama
  • NVIDIA GB10 / ASUS Ascent GX10
  • Nemotron
  • Phi-4
  • Purview
Is a small local model accurate enough?

Sometimes more so. In our bid-triage pipeline the 30B model found 11 fields the 120B model left empty and never fabricated an exclusion. We benchmark your task on byte-identical inputs before recommending a tier.

What hardware do we need?

For most document workloads, a workstation-class GPU or a $3–4k desktop box (we use an NVIDIA GB10 system) handles thousands of items a day. Laptops without an NPU can still run capable models for interactive use.

How is this governed?

Every prompt and answer passes through a safety gateway and is logged to Princeton Sentinel. Purview labels travel with the data; routing rules are versioned like code.

Can we start with one workflow?

Yes — that is how every engagement starts. We baseline it on cloud, move the suitable steps local, compare, and only then set routing and fallback rules.

Bring one workflow. We will show you where it should run.

A two-week hybrid AI assessment: baseline, local trial, routing recommendation, cost model.

Book the assessment