Rakshan
Anbu /

Failure forecasting for grid-edge assets. Ten-minute SCADA telemetry, weather reanalysis and work-order history land in a lakehouse, feed a gradient-boosted survival model, and surface as a ranked intervention list — each with the driver that pushed it up the queue, so an engineer can disagree with the model on the evidence.

A multi-agent research desk for regulatory change. A planner splits an incoming consultation or licence modification into questions; retrieval agents work a vectorised corpus of policy, contracts and internal controls; a critic agent hunts for unsupported claims before anything reaches a human. Every assertion carries a clause-level citation, and the run graph is replayable.

A resilience twin for a distribution network. Assets, feeders and customer connections are indexed on an H3 grid and layered against flood, subsidence and heat projections, so a planner can ask what a 1-in-100 event costs in customer-minutes lost and get an answer per substation rather than per region.

Near-real-time imbalance and settlement exposure. Half-hourly market data and metering flows stream through Event Hubs into Snowflake, get reconciled by dbt models with contract terms, and land on a trading-desk view that shows position drift while there is still time to act on it.

One semantic layer instead of forty conflicting definitions of "availability". Governed metrics are defined once in code, tested on every build, documented with lineage back to source, and exposed to self-serve users who can no longer accidentally invent their own denominator.

A connected production line. Edge cameras run a quantised detection model on-device to catch weld and seal defects at the station rather than at final inspection; MQTT carries cycle and downtime events upstream, where scrap is attributed to shift, tooling and material lot on the same page as margin.

Battery dispatch against a price forecast. A probabilistic day-ahead price model feeds a mixed-integer schedule that respects state of charge, degradation cost and grid service commitments, then a reinforcement-learning agent trades the intraday gap the deterministic plan leaves behind.