Six systems. Built, not pitched.

Five AI agents watch every plan at the same time. One reads the plan against the baseline. One matches what is happening now against what happened on past projects. One turns each risk into an action with an owner and a due date. One escalates when a threshold breaks. One writes the summary the board actually reads. Underneath, a dependency engine works out which slipped task drags everything else with it — before a human would have noticed.

What it proves

• Five agents running in parallel, each with its own job and its own limits — one failing does not take the others down
• A graph engine computing the critical path and cascading a single slip across every dependency
• Executive summaries generated from the same data the agents act on, so the board sees what the system sees

Watch the demoUnedited walkthrough with narration · Sherpa

One list of every outstanding invoice, ranked by risk. The AI drafts the chase email at three levels of firmness, so a first reminder does not read like a legal threat and a final notice does not read like a nudge. Disputes, promises to pay, part payments and handovers between collectors are all tracked in one place instead of in inboxes.

What it proves

• Tone-controlled drafting — professional, firm, escalation — with a human approving every send
• The AI only ever sees the accounts that collector is cleared for. Least privilege applied to the model, not just the app
• The whole collection lifecycle in one system: disputes, promises, part payments, statements of account

Watch the demoUnedited walkthrough with narration · Ark

The AI builds the plan in under a minute — but it is not allowed to invent anything. It picks only from 2,300+ meals the nutritionist has already approved, working through the client’s allergies, preferences and targets one step at a time. The expert stays the authority. The typing goes away.

What it proves

• Assembly, not generation — the model selects from an approved library, so it cannot make something up
• Eleven separate tools called in sequence, each doing one small job the expert can check
• Automatic failover between AI providers, so one provider going down does not stop the business

Watch the demoUnedited walkthrough with narration · NutriZen

Four layers that give an AI safe, useful access to what a company already has. It turns a database into a set of tools an agent can call. It runs two agents in sequence — one maps out what is in there, the other answers the question. It searches documents by meaning rather than keyword. And for the old systems with no way in, it drives the screen like a person would.

What it proves

• A database exposed as tools through MCP, usable by any AI agent, not locked to one vendor
• Two agents in an orchestrated sequence rather than one prompt asked to do everything
• Semantic search over a company knowledge base, and browser automation for systems with no API

Watch the demoUnedited walkthrough with narration · Mint

An assistant that answers on WhatsApp Business. Common requests — book a slot, check a time, send details — are recognised by pattern, with no AI call at all. That path answers in four milliseconds, costs nothing per message, and cannot get it wrong. AI is used only for the open-ended questions where it earns its place. It also transcribes voice notes and offers real appointment slots the customer taps to confirm.

What it proves

• Knowing when not to use AI — the deterministic path is faster, cheaper and more reliable for anything predictable
• Voice notes transcribed and acted on, not just stored
• Booking that ends in a confirmed slot in the calendar, not a promise to get back to you

Watch the demoUnedited walkthrough with narration · Elska

Sixteen database operations an AI can call — schema, health, statistics, backups, migrations, queries — with a safety layer in front of all of them. Anything destructive is refused unless a human explicitly confirms it in that same call. A small set of operations that can wipe out ownership across a whole database are never permitted, confirmation or not. This is what AI governance looks like when it is code instead of a policy document.

What it proves

• Twelve destructive patterns blocked by default — DROP TABLE, TRUNCATE, DELETE FROM and the rest — released only on explicit confirmation
• Two operations forbidden outright, with no confirmation available at any level
• The guardrails have their own test suite, so the safety layer is proven rather than assumed

execute_sql → SELECT count(*) FROM invoices WHERE status = ‘overdue’;
→ 1,284 rows · 41ms
execute_sql → DELETE FROM invoices WHERE status = ‘overdue’;
⚠ Destructive operation detected: "DELETE FROM". Pass confirm: true to execute.
execute_sql → DROP OWNED BY app_user;
✕ Forbidden operation: "DROP OWNED" is never allowed through pgman.
// the AI asked three times · the database survived all three