ServicesIndustriesWorkResolX AIAboutBook a consultation
Gen AI · Agentic AI · RAG since 2015

We build AI that survives contact with production.

Concept Box Technology has spent a decade shipping applied AI — multi-agent systems, retrieval platforms and computer vision — for healthcare, logistics, manufacturing and retail enterprises. We design the architecture, build the system, and stay until it runs.

NOVA AWS Challenge semifinalist Listed on the Deloitte Marketplace FDA submission workflows delivered
10+
Years building applied AI
From classical ML and computer vision through to agentic systems.
6
Regulated & heavy industries
Healthcare, life sciences, logistics, shipping, manufacturing, retail.
94%
Detection accuracy in production
YOLOv8 freight monitoring across US highway camera networks.
60%
Manual screening removed
AI hiring pipeline with voice interviews and semantic ranking.
AWS Bedrock AgentsLangChain / LangGraphGPT-4oClaudeLLaMA 70BSnowflakeYOLOv8Model Context ProtocolVector searchAWS LambdaFastAPINext.jsTensorFlowPower BIGroqElevenLabsQGISChromaAWS Bedrock AgentsLangChain / LangGraphGPT-4oClaudeLLaMA 70BSnowflakeYOLOv8Model Context ProtocolVector searchAWS LambdaFastAPINext.jsTensorFlowPower BIGroqElevenLabsQGISChroma

What we do

Eight capabilities. One team that has actually delivered each of them.

We are not a staffing shop with an AI page. Every capability below is backed by systems running in production for real clients in regulated and operationally demanding industries.

Advisory

The questions every board is asking. Here are our answers.

Most AI programmes fail on decisions made in the first month — the wrong model, the wrong infrastructure, an unexamined data posture. This is the conversation we have on day one.

Which model should we actually use?
We benchmark candidate models on your tasks with your data, and hand you quality, latency and cost per request side by side — frontier, mid-tier and open-weight.
What infrastructure do we need?
A right-sized architecture: serverless where load is spiky, dedicated capacity where it is not, and a clear picture of what each component costs at your projected volume.
Which cloud is right for us?
An objective comparison across AWS, Azure and GCP against your compliance regime, existing commitments, data residency and the models you need access to.
How do we keep customer data out of model training?
Zero-retention endpoints, contractual no-training terms, redaction before the boundary, tenant isolation and logged access — designed so you can prove it, not merely claim it.
How do we know it works before customers see it?
An evaluation harness with a golden dataset, regression gates in CI, and tracing on every production request, so quality is monitored rather than assumed.
What will this cost at scale?
A unit-economics model — cost per conversation, per document, per frame — with caching, routing and prompt design tuned to bring it down before launch.
See how an architecture engagement works

Selected work

Systems in production, not proofs of concept.

A sample of delivered engagements across agentic AI, generative AI and computer vision.

All case studies

Trust architecture

Your customers' data does not become someone else's training data.

The fastest way to lose an enterprise AI programme is to discover, after launch, that sensitive data crossed a boundary nobody mapped. We design that boundary first — and design it so you can prove where it sits.

  • Zero-retention model endpoints and contractual no-training terms, so customer data never enters a training set.
  • PII detection and redaction before anything crosses a model boundary.
  • Tenant isolation, scoped credentials and least-privilege tool access for every agent.
  • Full request tracing and audit logging — you can show a regulator exactly what happened and why.
  • Prompt-injection and jailbreak defences tested as part of delivery, not bolted on afterwards.
  • Deployment inside your own cloud account and region where data residency requires it.

How we work

Four steps, and an exit plan from the beginning.

Every engagement is designed so your team can eventually run the system without us. That is the point.

Step 01

Diagnose before we build

We start with the workflow, not the model. What does the work look like today, where does it actually cost you, and is AI the right instrument for it? Sometimes the honest answer is no, and we will say so.

Step 02

Architect and cost it

Model selection, cloud, data boundaries, guardrails and unit economics — decided on evidence and written down. You get an architecture your engineers can build against and a cost model your CFO can sign.

Step 03

Prove it on your data

A narrow, working slice against a golden evaluation set, on your real data. Measurable quality before anyone commits to a full programme.

Step 04

Ship, instrument, hand over

Production deployment with tracing, evaluation gates and cost monitoring — plus the documentation and training that lets your team own it. We are not trying to become permanent furniture.

Recognition

Semifinalist — NOVA AWS Challenge

AutoAdvisor multi-agent system

Nominee — Detroit Innovation Competition

AutoAdvisor multi-agent system

Listed on the Deloitte Marketplace

Validated enterprise readiness

Available for new engagements

Let us look at your use case properly.

A 30-minute call, no pitch deck. Bring the problem — you leave with a straight answer on feasibility, the architecture that fits, and what it would realistically cost to run.