$ what we ship

Six kinds of AI build.

We ship AI that has a job — deflecting tickets, summarising notes, triaging emails, drafting quotes. Every project has a metric, a baseline, and a definition of "worth the token cost."

01

RAG over your docsAssistants that answer from your policies, SOPs, product data — with citations. Vector search + reranking + guardrails to reduce hallucinations to near-zero.

02

Customer-support chatbotsOn-brand chat that deflects 30–50% of tickets before a human sees them. Hand-off flows, transcript export, and honest deflection reporting weekly.

03

Workflow automationn8n or Temporal workflows that read email, extract PDFs, sync CRMs, draft quotes and route approvals. LLMs where they add value, deterministic code everywhere else.

04

Agentic assistantsTool-using agents for internal ops — book appointments, run reports, update records, escalate exceptions. Sandbox-first, human-in-the-loop until the trace log earns trust.

05

Content & data pipelinesSummarisation, classification, extraction at scale. Batch jobs that turn 40 hours of manual review into 40 minutes of exception handling.

06

Model & provider strategyWhich LLM, which provider, self-hosted or API, prompt-caching strategy, cost per action. We benchmark on your data before we commit spend.

$ our process

From prompt to production — six phases.

Every engagement runs the same six phases. Fixed scope, fixed timeline, fixed price — agreed before we touch an API key.

01

Use-case & ROI framing
Week 1

Not every problem needs an LLM. We map the workflow, count the hours, pick the two or three places AI actually moves the number. If AI won't beat a script or a form, we say so.
- Workflow interviews with the people doing the task today
- Baseline metric: time, ticket cost, or error rate
- Model + provider shortlist with rough cost-per-action
- Fixed-price proposal with success metric written in
02

Data & prompt design
Week 1 – 3

Corpus prep for RAG (chunking, embeddings, source hygiene), prompt scaffolding with structured outputs, tool schemas for agent flows. Boring, meticulous — this is where quality lives.
- Corpus cleanup, dedup, canonical source of truth
- Embedding + reranking strategy (BM25 + dense)
- Structured outputs (JSON schema, tool calls)
- Prompt library versioned in Git, not Notion
03

Evals & guardrails
Week 3 – 5

A gold dataset of 100+ real cases, LLM-as-judge + human review, prompt-injection tests, hallucination checks. We measure quality before we ship it, and every prompt change re-runs the eval.
- Gold dataset from real historical tickets / docs
- LLM-as-judge scaffolding + human spot-checks
- Prompt-injection & jailbreak suite
- PII redaction + moderation classifier pass
04

Integration & UX
Week 5 – 8

Wire the model into where the work happens — Slack, Zendesk, HubSpot, your web app, n8n. Streaming responses, citations rendered, human hand-off flows tested end-to-end.
- Native integrations (Zendesk, Slack, HubSpot, Intercom)
- Streaming UI with citation rendering
- Human hand-off + escalation triggers
- Feedback capture (thumbs, correction) piped back to evals
05

Shadow-mode & pilot
Week 8 – 10

Ship to production in shadow mode first — model runs alongside humans without user impact. Two weeks of side-by-side data, then a real pilot with 10% of traffic.
- Shadow-mode telemetry: agreement rate, cost per call
- 10% pilot with A/B analytics
- Weekly review + prompt/model tuning
- Cost dashboards + budget alerts wired in
06

Launch & 60-day tuning
Week 10 → 18

Full rollout, on-call for 60 days for tuning. New failure modes get added to the eval set, prompts evolve, cost per action trends down. Weekly report showing hours saved vs token cost.
- Staged rollout with rollback if quality drops
- Weekly ops report: usage, cost, deflection, quality
- Prompt + eval updates for new failure modes
- 60 days of tuning & support — included

$ what you own

What lands in your hands.

Final milestone paid, everything transfers. Prompts, evals, API accounts, vector store — no lock-in, no rent-seeking.

Full source repo

Prompts, orchestration, tool schemas, evals — all in your GitHub org, versioned properly.

Model accounts & billing

Anthropic, OpenAI, Google, AWS Bedrock — all registered under your business, spend tracked to your card.

Vector store & corpus

Pinecone / pgvector / Weaviate — your tenant, your embeddings, your source docs re-ingestible any time.

Eval suite & gold dataset

Test cases, LLM-judge rubrics, historical results. Change a prompt, re-run the suite, ship with confidence.

Ops dashboards

Cost per action, latency percentiles, quality scores, deflection rate — Grafana or Metabase in your account.

Prompt & ops playbook

How to add a new source, how to tweak a prompt, how to handle a model provider outage — written for humans.

$ tech stack

The stack we reach for.

Boring on purpose. Every tool here is battle-tested, actively maintained, and hire-able for anywhere in Brisbane, Sydney, Melbourne, Auckland or Wellington if you want to bring it in-house later.

Model providers

Anthropic (Claude)
OpenAI (GPT)
Google (Gemini)
AWS Bedrock
Azure OpenAI

Orchestration

Claude Agent SDK
LangGraph
Vercel AI SDK
Temporal

Retrieval

pgvector
Pinecone
Weaviate
BM25 hybrid
Cohere Rerank

Automation

n8n
Zapier
Temporal
Airflow

Evals & observability

Braintrust
LangSmith
Helicone
Langfuse

Runtime

Node.js
Python
FastAPI
Cloudflare Workers

$ timeline & investment

Rough shape of an AI engagement.

Real numbers from the last dozen AI builds. Actual pricing lands after discovery — but this is the sensible band to plan around.

Typical timeline

0–0 weeks

Focused chatbot or RAG assistant: 3–6 weeks. Multi-tool agent with evals: 8–10 weeks. Full workflow automation with human hand-off: 10–14 weeks.

Investment (AUD)

$0k – $0k

Fixed price, three milestone payments. Token cost passed through at cost. Care plan from $380/mo covers eval maintenance + model version bumps.

0

LLM projects shipped

0M

Tokens processed / month

0 min

Avg first reply

0%

Client retention

$ questions

The stuff every client asks first.

If your question isn't below, we'll answer it on the discovery call. No sales pitch — just an honest read on whether AI actually fits.

Do I actually need AI or will a form / script do?

Honestly, about a third of "we need AI" conversations end with us recommending a form, a Zap, or a small script that costs a tenth of an LLM build. AI earns its cost when the task involves fuzzy natural language, unstructured docs, or judgement. We'll tell you which bucket you're in.

Which LLM should I use — Claude, GPT, Gemini?

Depends on the task. Claude is our default for long-context RAG and agentic tool use. GPT-4o and o3-class models shine for reasoning-heavy or structured extraction. Gemini for very long context + native multimodal + tight GCP integration. We benchmark on your data before we commit spend — the answer isn't a religion.

Will it hallucinate on our customers?

Not zero, but close. RAG + citations + reranking + structured outputs + eval gates knock hallucination rates well below human error rates on well-scoped tasks. For high-stakes flows we keep humans in the loop until the trace log earns trust. We report failure modes, not marketing numbers.

What about our data — is it safe with the model providers?

Anthropic, OpenAI Enterprise, and Google Vertex don't train on API data. For sensitive workloads we use AWS Bedrock or Azure OpenAI in Sydney region, or self-host open models on your infra. We build to your data-residency and privacy requirements, not the other way around.

What does ongoing token cost look like?

Depends on volume. Typical support chatbot: AUD $200–$800/mo in token cost. RAG assistant for a mid-sized team: $400–$2k. High-volume batch pipelines: $2k–$10k. We wire cost dashboards + budget alerts on day one so surprises don't happen.

Do we use n8n or custom code?

n8n for automation workflows with lots of integrations and non-dev editors. Custom code for anything mission-critical, high-volume, or that needs proper evals + version control. Often we ship a hybrid — n8n for the plumbing, TypeScript/Python service for the LLM call itself.

Do you work outside Brisbane?

Yes — every engagement runs remote-first from our Brisbane studio, with on-site availability across Sydney, Melbourne, Adelaide, Perth, Canberra, Auckland and Wellington. AI kickoff workshops work particularly well in person; most implementation runs on Slack + Zoom.

AI that pays for itself in month one.