MalvaOS doesn't replace anything. It sits on top, and each agent pulls from the tools below.

The operator keeps the software they already trust. MalvaOS reads from those systems, lets the agents reason across them together, and writes back as recommendations — never a forced action.

Operator's existing stackOptional third-party specialist

◇

Geology

Owns the block model and ore-grade calls.

reads from →LeapfrogDrillhole DBSatellite imagery+Mineflow+KoBold AI+Plotlogic

✺

Blasting

Designs the drill pattern and powder factor for each bench.

reads from →Blast design SWVibration monitorsGeotech sensors+Strayos+BlastIQ

◐

Trucks & Diggers

Dispatches excavators and haul trucks for best $/cycle.

reads from →Wenco / Modular FMSGPS telemetryPayload sensors+Newtrax+Pitram

◼

Crusher & Bag Line

Keeps the 30 t/h comminution circuit and bag line on spec.

reads from →Citect / AVEVA PILIMSBag-line PLC+MineSense+Petra Data Science

◯

Maintenance

Predicts failures and books work orders before things break.

reads from →Pronto / SAP PMVibration sensorsOil analysis+Uptake+SparkCognition

▲

Logistics

Keeps diesel, consumables and outbound concentrate moving.

reads from →ERPFuel telemetryWeather feeds+Convoy tracking APIs+Sixfold / project44

Why this compounds

Nothing gets ripped out.

Existing systems stay in place. MalvaOS reads from them through APIs and exports — it doesn't try to replace any vendor.

Data stops being siloed.

One question can pull answers from FMS + SCADA + LIMS at once. That cross-system reasoning is where the gains live.

Every new tool makes every agent smarter.

Add Mineflow → Geology sharpens. Add Strayos → Blast sharpens. Plant and Fleet inherit the better data automatically.

How it learns

In the demo the agents follow scripted scenarios. In production every recommendation is logged with its actual outcome (head grade, throughput, downtime, $/t) and fed back into a reward model. The orchestrator learns which agent combinations and parameter choices actually move the dollar metric — not just the theoretical one.

Before — Week 1

Naive dispatch

Fleet dispatches trucks to the nearest high-grade stockpile by GPS distance alone. Head grade lands inside spec 72% of shifts.

Head grade on spec: 72% · Bags/shift: 840

After — Week 6

Dispatch + learned blend policy

After ~200 shifts of feedback, the orchestrator couples Fleet with Geology. Dispatch is conditioned on the block model — it knows which face blends with which ROM stockpile to hold the 2.0% Sb target.

Head grade on spec: 94% · Bags/shift: 1,020

📈

ROI from the data flywheel: +$312,000 over six weeks from fewer off-spec batches and steadier bag-line output. The more shifts it runs, the tighter the predictions get.

Vendor-neutral by design. The orchestrator's value grows with every system it connects to.