Awards & Recognitions
Generative AI development that holds up in production
We build chat over your documents, AI assistants for your teams and LLM features inside your product, grounded in your content, tested before launch and connected to your systems.
Can contractors claim travel for site visits?
Yes, if the visit is in their statement of work and booked through the travel portal
[1]. Mileage is paid at the standard rate; flights need manager approval
first [2].
[1]Contractor handbook · section 4.2
[2]Travel and expenses policy · page 3
Answered from 2 approved sources
Access checked
Under 800ms
responses in a real-time AI interview platform
Ignyte AI
retrieval-based recruitment platform for KingsleyGate
2 AI products
of our own in production: VoAgents and WellBeing Navigator
Since 2012
CMMI Level 3 · ISO 27001
What we build
Generative AI services we deliver
Every build includes the retrieval, guardrails, testing and integration it needs to go live. Pick what you want to build to see the stack we’d put together, and what we’d decide with you first.
1. What do you want to build?
2. What does it need to do?
3. Model
4. Hosting
Chat over company documents
7 of 10 layersThe stack Orion would build. Faded layers aren’t needed for this setup.
Interface
Chat in your intranet or web app
Assistant inside the tools your team uses
Chat widget on your website or app
Feature in your product UI, plus an API
Orchestration
LangGraph agent that plans steps and calls tools
LangChain pipeline: retrieve, prompt, check, answer
Model
OpenAI models, with a fallback configured
Claude models, with a fallback configured
OpenAI or Claude, chosen after testing on your content
Retrieval
Your documents indexed and searched; every answer cites its source
passages
Not needed for this setup
Actions
Tools that read and update your systems through their APIs, with
permission checks
Not needed for this setup
Guardrails
Answers limited to approved content; off-topic and unsafe requests
refused
Access checks, sensitive-data masking and policy rules
Human handoff
Hands the conversation to a person when it can’t answer
confidently
Not needed for this setup
Tenancy
Each customer’s data kept separate, with per-tenant access
control
Not needed for this setup
Evaluation
A test set of real questions scored before launch and after every change
Hosting
Azure, North America, with usage and cost monitoring
AWS, North America, with usage and cost monitoring
We’d decide with you first
- Which documents are in scope, and who may see what
- Which product data the feature may use
- Whether OpenAI meets your accuracy and cost targets
- Whether Claude meets your accuracy and cost targets
- Which model performs best on your own test questions
- Which actions it may take on its own, and which need approval
- What it should refuse to answer
- What a good answer looks like, written as a test set
- When to hand over to a person, and to whom
- How tenants are separated and billed for usage
Production-grade
What separates a demo from production
Same question, same documents. The difference is what we build around the model.
A typical demo
No sources
What’s our refund window for annual plans?
Annual plans can be refunded within 60 days of purchase, no questions asked.
- States a policy that doesn’t exist
- No source, so nobody can check it
- Answers staff who shouldn’t see it
Built by Orion
Cited · checked
What’s our refund window for annual plans?
Annual plans can be refunded within 30 days of purchase [1]. After that, unused months are credited, not refunded [2].
Access checked · Finance team
- Answer taken from your current policy
- Sources cited, so anyone can check
- Access checked against your permissions
Illustrative example.
How it works
How a grounded answer is produced
Six steps sit between your documents and the answer your users see. A test set built from your own questions measures the whole chain.
-
01
Your content
Documents, policies and data you choose to include.
-
02
Indexed
Split into passages and indexed, keeping each one’s access rules.
-
03
Retrieved
The passages that answer the question are found, filtered by who’s asking.
-
04
Model drafts
OpenAI or Claude writes the answer from those passages only.
-
05
Checked
Guardrails check access, sources and policy before anything is shown.
-
06
Answered
The user gets the answer with sources cited, and the exchange is logged.
Tested on your own questions before launch and after every change, so accuracy is
measured, not assumed.
Grounded in your content
Answers come from your documents and data, with the source shown.
Tested before launch
Scored on real questions, so you know how often it’s right before users see it.
Guardrails
Limits on what it can say and do, with sensitive data masked.
A person when needed
Uncertain or sensitive conversations go to your team.
Cost and speed watched
Usage, response time and spend tracked from day one.
You own it
Code, prompts and test sets are handed over. No lock-in.
Models: OpenAI · Claude
Orchestration: LangChain · LangGraph
Hosting: Azure · AWS, North America
Process
From prototype to production in five steps
You see it working on your own content early. Orange marks the point where you decide whether it’s ready.
-
01
Scope
Agree the users, the content and what a good answer looks like.
You get: A written scope and a starter test set. -
02
Prototype
A working version on a sample of your own content.
You get: Something your team can try. -
03
Evaluate
Score it on real questions for accuracy, safety, speed and cost.
You get: Results and a go or no-go decision. -
04
Harden
Add guardrails, access control, integrations and monitoring.
You get: A production-ready build. -
05
Launch and improve
Roll out, watch quality and cost, and retest after every change.
You get: Monitoring and a regular quality report.
Case studies
Generative AI we’ve shipped
For clients, and for our own products.
KingsleyGate
AI recruitment platform with retrieval at enterprise scale
A multi-tenant recruitment platform where AI matches candidates to roles, evaluates competencies using retrieval over candidate records, and analyses interview transcripts.
Service: Custom LLM app
Steps used in this build
- 1Content
- 2Index
- 3Retrieve
- 4Draft
- 5Check
- 6Answer
What we built in
Retrieval-based competency evaluation
Transcript analysis
Strict data isolation between tenants
Role-based access control
Orion product
Our own conversational AI coach
A wellness product that gives personalised, on-demand coaching through natural conversation and adapts to each user over time.
Service: AI assistant
Steps used in this build
- 1Content
- 2Index
- 3Retrieve
- 4Draft
- 5Check
- 6Answer
What we built in
Conversational coaching
Personalisation over time
Privacy-first: user data is never sold
Missional Software
A receptionist that answers only from approved content
A multi-tenant assistant that detects what a caller needs and responds only from verified content, routing urgent requests to staff straight away.
Service: Customer-facing assistant
Steps used in this build
- 1Content
- 2Index
- 3Retrieve
- 4Draft
- 5Check
- 6Answer
What we built in
Answers limited to approved content
Guardrails against made-up answers
Urgent requests routed to staff
Multi-tenant
Frequently asked
Questions buyers ask first
About accuracy, data and what you own.
Ask us something elseAnswers are generated from your own content through retrieval and must cite their sources. If nothing relevant is found, the assistant says so or hands over to a person instead of guessing.
We build on OpenAI and Claude models, orchestrated with LangChain or LangGraph. We test the candidates on your own questions and pick on accuracy, speed and cost.
In your Azure or AWS environment in North America, or in ours if you prefer. Your content isn’t used to train public models.
We build a test set of real questions with agreed answers and score every version against it before launch and after each change, alongside response time and cost.
Yes. Assistants can read from and act in your CRM, ERP, helpdesk or document stores through their APIs, with permission checks on every action.
You do. Code, prompts, test sets and documentation are handed over, whether or not Orion runs it afterwards.
Let’s talk
Tell us what you want to build
A GenAI engineer replies within one business day with questions, a suggested approach, and an honest view on whether an LLM is the right tool.
NDA available on request.
Send us your enquiry
A GenAI engineer will reply within one business day.
NDA available on request.
