The nine specialist agents.
ParallaxOS AI is a multi-agent system. The Auto agent routes every plain-English query to the correct specialist, and each specialist is an expert in one part of the rail subcontractor business. Powered by the Anthropic Claude API.
How routing works
You don't pick an agent. You ask a question, and the Auto agent classifies it against the specialist domains, then forwards to the right one. The user sees the answer with the agent name attached for transparency.
// Example query lifecycle user: "Can Mike Davies work ARTC tonight?" auto: "→ Compliance agent (worker readiness query)" comp: "Yes. Mike's RIW card is valid until 2027-04-15, his Cat 1 medical is current to 2026-08-22, and his FAID score is 55/100 (rosterable)."
The Auto agent
The router. Single responsibility: classify the query and forward it. Returns the specialist's answer verbatim, prefixed with which agent handled it.
The eight specialists
1 · Compliance agent
Knows every worker's exact certification status across every supported network. Answers questions like:
- "Who can work TfNSW tomorrow?"
- "How many of my workers have a current Cat 1 medical?"
- "What's blocking Asha Patel right now?"
2 · Documents agent
Indexes every uploaded SWMS, EBA, ARTC Network Rule, and network notice. Cited answers in two seconds.
- "What does section 4.2 of the J001 SWMS require?"
- "Where is the emergency evacuation point on the Muswellbrook job?"
- "Show me the chain-of-custody clause from our SWMS."
3 · ESG agent
Generates monthly GHG reports with fuel breakdown by asset, Scope 1 and 2 calculations, carbon intensity per hour, and YTD totals — formatted for Avetta ESG submission.
- "What's our Feb 2026 GHG total?"
- "Project Scope 1 emissions for the rest of FY26."
- "Generate the AASB S2 climate disclosure for last quarter."
4 · Awards agent
Reads your Enterprise Agreement and answers penalty-rate questions with the exact clause and dollar amount cited.
- "What penalty rate applies to a Sunday night rail shift for a Safeworking Officer?" → "$153/hr — Clause 14.2(b) of the Enterprise Agreement (base $85 × 1.8 weekend penalty)."
- "What's the meal allowance for a 14-hour shift?"
5 · Fatigue agent
Triggers automatically when a worker's FAID score exceeds 75/100. Alerts the supervisor by SMS, offers to dispatch an Uber, and seals the FAID event for the compliance record.
Worked example from the Hunter Valley reference dataset: Mike Davies finishes an 11h 28m shift at Muswellbrook with a FAID of 78/100. Rest period 9.5h, below the 10-hour minimum. Supervisor Ben O'Connor receives an SMS alert instantly. Push notification to Mike: "High fatigue risk — please rest before next shift." Uber dispatched: Driver Raj, 7-min ETA, Muswellbrook Yard. The FAID event is logged and added to the compliance record.
6 · Training agent
Identifies workers with expiring competencies and recommends nearby RIW-approved RTOs with course dates and GPS distance.
- "Who needs recertification in the next 60 days?"
- "Find an Aurizon-recognised SWO refresher course near Muswellbrook this month."
7 · Medical agent
Finds GPS-sorted Authorised Health Professionals for any worker needing a medical, showing Cat 1/2/3 capability and live RIW Direct Upload status.
- "Book Steve Martin's Cat 1 medical at the closest RIW-Direct AHP."
8 · Pre-Start agent
Guides workers through their pre-start checklist on mobile. When connectivity drops, a Phi-3 Mini on-device model handles basic queries until signal returns.
Conversation history
Conversations persist with full search on Professional and Enterprise — review what the ESG agent said on 3 February, or export an Awards-agent calculation as a PDF for a compliance record.
Custom training (Enterprise)
Upload your specific EBAs, ARTC network notices, client contracts, and internal SWMS templates. The platform indexes them so subsequent queries pull from your actual documents. Significantly improves accuracy on domain-specific questions.
What the AI doesn't do
The agents do not make assignment decisions. The compliance gate is rule-based and deterministic — workers are blocked or cleared by hard rules, not by the AI's judgement. The AI surfaces, explains, and recommends; the platform enforces.
Voice input on mobile
Workers in the field can use voice-to-text with the Pre-Start and Fatigue agents — typing in hi-vis gloves at 04:00 in a tunnel is not realistic.