AI in production
AI your team can trust after the demo.
Almost every software firm now calls itself AI-native. Far fewer can tell you how often their agent is wrong, what each task costs, or who approved the change it made. That part — the measurement, the limits and the approvals — is what I build.
Measured before launch
You know the error rate before your customers do.
- Test sets built from your real cases, including the awkward ones
- Error rates by task type, tracked release to release
- Regression checks before a prompt or model change goes live
- A clear threshold for when the feature is good enough to ship
Costs with a ceiling
Every request has a budget, and someone sees the bill.
- Per-task and per-day spend limits that stop work, not just warn
- The smallest model that passes the test set, not the biggest available
- Caching and batching where the answers allow it
- Cost per task reported next to quality
A person approves anything irreversible
AI drafts. A named person confirms.
- Sending, paying, deleting and posting always wait for a person
- Drafts checked by your own system's rules before anyone sees them
- An audit trail: what was proposed, what changed, who approved it
- A one-click way to turn the feature off
Safe tool access for agents
An agent can only do what its job needs.
- Explicit tools with the narrowest permissions that work
- Defences against prompt injection from emails, documents and web pages
- MCP servers and API keys scoped, rotated and logged
- Stop conditions on steps, time and spend
Data boundaries & DPDP
You decide what leaves your systems.
- Personal data mapped before a single prompt is written
- Grounding on your own data, respecting each user's permissions
- What is logged, where, and for how long, written down
- Designed for India's DPDP Act, and GDPR where you sell to Europe
AI inside ERPNext & Frappe
The same rules, inside the system that runs your business.
- Claude running inside Frappe on its own worker queue
- Your own key, stored write-only, one session at a time
- Leads to briefs, opportunities to quotations, projects and tasks
- ERPNext prices and validates every draft; nothing is written without approval
Proof
I built this for my own business first.
My practice runs on ERPNext with an AI layer I wrote: Claude reads a new enquiry and drafts a brief; reads an opportunity and drafts the quotation; and, once the client approves, drafts the sales order, project and tasks.
Claude never writes to ERPNext directly. Every proposal is priced and dry-run by ERPNext first, shown as ready only if it passes, and created only when I confirm it. That's the pattern I bring to your systems.
Services in this area
AI readiness review
One workflow mapped, measured and given a go or no-go.
Proof of concept
A time-boxed build on one workflow, measured against a test set.
Build to production
A signed-off design delivered with evaluations, spend limits and approvals.
AI feature development
Changes to prompts, tools and integrations in a feature already live.
Evaluation & cost review
An independent look at quality, cost per task and risk in an existing feature.
Fractional AI architect
Design, prompt and release reviews on a reserved block of hours.
Next step