AI in Action 2

Proof of capability · Applied AI in production

We don’t pitch AI deployment. We run it in production.

Most AI agency decks are slideware. Ours is a live platform, real users, run by our own AI operator Atlas, with a person on every call that matters.

StudyVisaHubLive production platform
AtlasAgentic engineering operator
GovernedConstitution + authority matrix
DataOpenly-licensed & official only
Atlas · this cycleSupervised
Selected next task from backlogauto
Build gate: lint · type-check · buildpassed
Independent verify circuitpassed
Merge to productionhuman approves
Class 2 actions never auto-run · human on data & money
01 · Why this document exists

The hard part of AI isn’t the model. It’s trusting it in production.

Everyone can buy the same AI. The gap between a demo and a real deployment is everything around it: messy data, real workflows, proof it behaves. That’s the work we already do.

~95%
of enterprise GenAI pilots show no measurable business impact (MIT, 2025)
Integration
the cause, not the model. Messy data meeting real workflows
Live
our platform runs in production on regulated data today
Human-gated
every irreversible action stops for a person, by design
What this is, and isn’t

Supervised autonomy, not a hands-off robot. The AI does the engineering; clear rules decide what stops for a person; anything risky is human-approved. That gate is why it can be trusted to run at all.

02 · The platform

StudyVisaHub: a live product built on real, regulated data.

A production platform for international students, education providers and migration agents. Not a sandbox. Real users, on data that has to be correct.

THE DATA
Regulated & unforgiving
Official visa grant rates, country evidence levels, accredited provider and course records, scholarships and PR-pathway intelligence. Wrong data isn’t cosmetic, it erodes trust.
THE SCOPE
Six defined domains
Visa Data · Providers · Agents · Courses · Scholarships · Career & PR. Bounded on purpose, the operator can’t invent features outside these domains without sign-off.
THE STACK
Production-grade
Next.js, React, TypeScript, PostgreSQL, cloud hosting, CI. A real deployment pipeline with lint, type-check, build and CI gates, not a prototype.
03 · How the data is sourced

Openly-licensed, official, and downloaded from the source.

Core datasets are official government data, openly licensed (CC BY) and pulled straight from data.gov.au, with the source credited.

Official source data.gov.au DHA · Education Licence: CC BY attribution recorded Downloaded direct from data.gov.au Validation pipeline: clean, check, structure Human QA then published with attribution Creative Commons Attribution (CC BY). Core datasets (visa statistics & CRICOS) are open Australian government data from data.gov.au, credited to the publisher.
Every dataset enters through the same audited chain, provenance and licence recorded before it reaches a user.
Why this matters to you

Provenance is a liability question, not just a quality one. The same discipline points at your project: we use only data you own or are licensed to use, with the source and licence recorded.

The operating principle

Mostly deterministic software. AI only at the genuine judgment points. A human on anything irreversible. That ratio is what makes the system trustworthy, and cheap to run.

04 · The agentic operator

Atlas: an agentic AI engineer that ships real work, supervised.

Atlas picks the next task, writes it, and checks its own work. A person approves anything that goes live. Routine engineering, off your plate.

Every change is small, checked automatically, and reviewed by a person before it ships. The discipline is the point:

SCOPE
One bounded task
Each change is a single, clearly-scoped piece of work (a fix, a refactor, a test, an internal tool) never a sprawling compound edit that’s hard to review.
GATE
Full build gate
Every change must pass lint, type-check and build before it can enter the review queue. No green build, no queue, enforced in code.
MERGE
Human merges
Atlas opens the pull request; a person reviews and merges. It never merges its own work, and never touches production data or schema on its own.
The pattern that makes it safe

The AI does the build. Anything that touches the database or live data is drafted for a person to approve, never applied on its own. That one boundary is the difference between a useful operator and an expensive risk.

05 · The governance model

Knowing where autonomy stops, decided in advance, enforced in code.

The clever part isn’t making an AI do more. It’s deciding in advance what it may do alone, what it drafts for a person, and what it must never touch, enforced by software.

Task arrives Classify the action Class 1 · Recoverable bugs, refactors, tests, PRs Product / business impact behaviour or scope change Class 2 · Irreversible data, secrets, spend, billing AUTO agent builds & opens a PR ≈ 90% of the work PROPOSE drafted for a human not actioned HUMAN ONLY agent never acts surfaces & moves on
Every task is classified before it runs. The verdict (not the agent’s judgment in the moment) decides who acts.

Hard stops the operator cannot talk its way past

Enforced in code. The AI can’t grant itself more power or talk its way past a no, even if asked.

HARD STOP
Never self-merges
Opening a PR is as far as automation goes. Merging is a human decision.
HARD STOP
Never migrates data
Schema changes are proposed as their own PR for a person to apply.
HARD STOP
Never touches secrets
Environment keys, auth rules, payment webhooks, source inputs are off-limits.
HARD STOP
Never edits its rules
Governance files are fingerprinted; if they change, work halts for re-verification.
Two-speed by design, match the control to the risk

Small, reversible changes ship fast. Anything touching the data goes through the full check before it’s trusted. Speed where it’s cheap, rigour where it counts.

06 · Evidence, not vibes

We never trust the agent’s own word for done.

The most common failure of AI agents is grading their own homework. We designed against that from the start.

Agent reports “done”, a claim Independent verify circuit separate code, re-checks from scratch working tree lint type-check build diff-scope secret scan every check must pass, no partial credit all pass? yes / no pass Queued for human merge any single check fails → NOT done, returned to the agent
The completion report is treated as a claim to be tested, never as proof.
FAILURE MODE WE GUARD AGAINST
Self-congratulation
An agent’s “I finished, it works” is not independent evidence. Our countermeasure re-runs the checks itself and reads the result, not the summary line.
FAILURE MODE WE GUARD AGAINST
Silent drift
Self-improving agents can quietly erase a safety constraint a human set. We keep the rules where the agent can’t edit them, fingerprinted, so any change stops the line.
written constitution two-tier authority matrix independent verify circuit no partial credit human gate on money & data full audit trail fingerprinted guardrails
07 · What this means for your business

The same method, pointed at your workflow.

We built this running our own platform. Pointed at your business, it’s the same method, aimed at a back-office job costing you time and errors today.

The worked example, our own production pipeline

The exact chain behind every change: sorted by risk, built, then independently checked before a person approves it. AI for judgment, plain code for fixed rules, a person on anything irreversible.

Human decision Class 2 or business impact → surfaced, never auto-actioned Backlog queued work Select next task Classify governance → AUTO? Build the change Verify circuit Human merges → live not AUTO verify fails → rebuild
AI judgment Deterministic rule Human gate Input
The judgment lives with the model; the rules live in code; the merge decision lives with a person. This runs every task on the platform, today.

How every deployment is built, four stages

1
Sprint design

Map how the work actually happens, including every exception the SOP doesn’t mention.

2
Judgment map

Classify each step: deterministic rule, AI judgment, or human gate, the same owner split shown in the pipeline above.

3
Evals

Prove it behaves on real cases before it goes near production, evidence, not a demo.

4
Graduated deploy

Shadow → human on every action → widening autonomy, never past the risk gate.

The honest version, because that’s the whole point

No magic box that runs your company unattended. Something better: we put AI to work on real tasks safely, prove it with evidence, keep a person on the big calls, and show you a live system where we already do it. Most people selling AI can’t.

The bottom line

We can show you a live system where we already do this, safely, supervised, with the evidence to prove it. Most people selling AI deployment can’t.

Next step

See the method pointed at your business.

Book a diagnostic and we’ll map the highest-value workflow to automate first, with the governance, evidence and human gates built in from day one.