Solution · 04 · CST
· BESPOKE / MRR
Custom AI & Secure Chatbots.
Private copilots and customer-facing assistants built on your data — RAG systems hardened against prompt injection, jailbreaks, and data exfiltration. So you can actually ship them to production, not just demo them.
What we build
— 01
Three shapes of custom AI.
Private copilots for your team, customer-facing assistants, and models trained on your data. Same underlying stack, different target users.
Shape · 01 · Internal
Private copilots
Slack-native, browser-extension, or embedded in your app. Answers from your SharePoint, Notion, Confluence, tickets, and code. Never trained on public data.
Shape · 02 · Customer
Support & sales bots
Web widget or WhatsApp. Handles tier-1 support, qualifies leads, escalates cleanly. Tuned for your brand voice, bounded to your domain.
Shape · 03 · Custom models
Fine-tunes & RAG
When a generic model isn't enough — fine-tuned classifiers, domain extraction, proprietary embedding spaces. With evals to prove they beat the baseline.
Hardening
— 02
Security isn't a feature — it's the deliverable.
Every bot we ship goes through the same five-layer hardening pass. This is what separates a production system from a demo.
Layer · 01
Input guardrails
Every user message is screened against known prompt-injection patterns, jailbreak templates, and out-of-scope requests. Malicious inputs are rejected before they reach the model.
PROMPT INJECTION
Layer · 02
Policy enforcement
System prompts, tool permissions, and data-access policies are declared as code. The model can only see and do what policy allows — not what the prompt says.
SCOPE CONTROL
Layer · 03
Output filtering
PII, secrets, and internal codenames are stripped before responses leave the server. Retrieved context is redacted at the chunk level, not hoped-for at generation time.
DATA LEAK
Layer · 04
Full audit log
Every request, retrieval, tool call, and response is logged with user, timestamp, and trace ID. Searchable dashboards. Retention configurable for compliance.
TRACEABILITY
Layer · 05
Quarterly red teaming
We attack the bot every quarter with the latest jailbreaks, data-exfiltration patterns, and social-engineering attempts. You get the report. Zero critical findings in 12 months.
ONGOING
Use cases
— 03
Production deployments, anonymised.
1
Internal knowledge copilot
"How do we do X?" answered from your wiki, in Slack, in seconds
Internal
2
Tier-1 support bot
Handles 60–70% of incoming tickets without human review
Customer
3
Onboarding assistant
Trained on your playbooks — the first week of a new hire, compressed
Internal
4
Sales copilot
Real-time pitch support grounded in your playbook & pricing
Revenue
5
Compliance bot
Policy Q&A with citations — legal & HR teams love this one
Internal
6
Technical docs assistant
Embedded in your product; grounded on code & docs, not web
Customer
Questions
— 04
What security-conscious teams ask first.
If your CISO has the question, we've had it from someone's CISO before. Happy to talk in detail.
Does my data ever train a third-party model?
No. We run on enterprise agreements with OpenAI and Anthropic that explicitly exclude training. Full DPA available. Data lives in your region.
Q·01
How do you handle prompt injection?
Multiple layers. Input classification, tool-permission whitelisting, sandboxed retrieval, and output filtering. Every bot is red-teamed before launch and every quarter after.
Q·02
Can we run it fully on-prem?
Yes. For regulated customers we deploy on your infra with open-weight models (Llama, Mistral) and self-hosted vector DBs. Performance gap is small, compliance story is airtight.
Q·03
Who owns the bot after launch?
You. Prompts, tools, and data pipelines are yours, in your git. We run it on MRR; you can take it in-house any time with full documentation.
Q·04
More solutions
— keep exploring
Related ways we ship AI.
Ship an AI your CISO trusts.
30-minute call. Bring your data story, compliance constraints, and the use case. We'll tell you if it's a fit, what the build looks like, and how the hardening works.
Email
Office
Melbourne, VIC · Remote available
Response
< 24 hours · always