AI Development & Integration in New Zealand
AI features that earn their keep, integrated into your product or workflow and proven first in our own: reading meal photos, tailoring CVs and matching receipts to bank transactions, live in production today.
Most AI development advice comes from people who have shipped a demo, not a production feature carrying real users and real cost. Corvidae runs generative AI, computer vision, OCR, retrieval-augmented generation and MCP agent servers in our own live products every day, so we know where AI genuinely earns its keep and where it is a feature for its own sake.
We build AI development and integration for NZ businesses on that same experience: grounded in your own data, evaluated against real cases, and priced to control the ongoing cost of running a model in production.
Who it is for
- Businesses with a specific, well-defined problem AI could solve, not a vague mandate to add AI
- Teams that need AI grounded in their own documents and data, not generic chat
- Products that need document, image or receipt processing at volume
- Organisations wanting an AI agent or MCP server that can safely call their own systems
- Anyone who has been burned by an AI feature that looked good in a demo and fell over in production
Problems we solve
- Manual document, receipt or form processing that does not scale
- Support and admin tasks that eat hours a week doing something an LLM can draft
- Knowledge locked in documents nobody has time to search
- AI pilots that never made it past a proof of concept
- Uncontrolled AI spend with no visibility into cost per request
What we build
Generative AI
- drafting
- summarisation
- classification
- extraction
- content tailoring
RAG and grounded AI
- document ingestion
- embeddings
- retrieval
- citations
- private knowledge bases
Vision and OCR
- document extraction
- receipt processing
- image understanding
- structured data extraction
AI agents and MCP
Our own MCP servers already expose 107 AI tools live to AI assistants such as Claude and ChatGPT.
- MCP servers
- agent tools
- workflow integration
- authentication
- permissions
- observability
Evaluation and cost control
AI that ships without evaluation is a liability, not a feature.
- model evaluation
- hallucination testing
- cost monitoring
- latency tracking
- prompt and version tracking
- fallback models
Technology
How we deliver
Assess
We separate a genuine AI opportunity from a nice-to-have, with a written scope before any model gets called.
Prototype
A working proof of concept against your real data, evaluated on accuracy and cost before we commit to a build.
Build
Production-grade integration: retries, fallbacks, logging and cost controls, not a bare API call.
Operate
We monitor accuracy, latency and spend once it is live, and keep tuning it.
Production proof
SnapKai reads meal photos, RoleRamp tailors CVs and Moniaro Books matches receipts, all live today.
Corvidae builds and operates its own production SaaS products, so architecture decisions are made with real operating cost, uptime, security, support and users in mind.
Security and quality
- Company data stays in your own cloud environment; we do not send it to shared third-party training pipelines
- Prompts, models and outputs are versioned so behaviour changes are traceable
- Cost and latency are monitored per request, not discovered on the invoice
- Fallback paths for when a model is unavailable or returns a low-confidence result
- The same senior engineer designs and reviews every AI integration we ship
Engagement model
AI engagements usually start with a short feasibility assessment: is this a good AI problem, what would it cost to run, and what does success look like. From there we build a working prototype against your real data before committing to a full integration.
Ongoing engagements can include monitoring and tuning cost, accuracy and latency once the feature is live, since production AI needs to be watched, not shipped and forgotten.
Frequently asked questions
Can AI be integrated into an existing system?
Yes. Most of our AI work is added to a system that already exists, whether that is document processing bolted onto an existing app or a chat feature grounded in an existing knowledge base.
Can company data remain private?
Yes. We build AI features that run inside your own cloud environment and call model providers directly, so your data is not pooled with other customers or used to train shared models.
What is RAG?
Retrieval-augmented generation. Instead of relying on a model's general training, we retrieve the relevant passages from your own documents first, then have the model answer grounded in those passages, with citations back to the source.
Can you build an AI agent?
Yes. We build agent tools and MCP servers that let an AI assistant call your product directly, with the same authentication and permission boundaries as a human user.
Which AI models do you use?
We build primarily on AWS Bedrock and Claude, and choose vision or language models case by case based on accuracy, latency and cost for the specific task, rather than defaulting to one model for everything.
How do you control AI costs?
By monitoring cost per request from day one, choosing the smallest model that meets the accuracy bar, caching where it makes sense, and setting fallbacks and limits so a spike in usage does not turn into a spike in the bill.
How do you test AI output quality?
With an evaluation set of real or realistic cases specific to your use case, checked for accuracy and hallucination before launch and monitored on an ongoing basis after it.
Have a product in mind?
Tell us what you want to build. We reply within one business day with an honest read on scope, cost and timeline.
Start the conversation