Power Procurement
Optimisation
Agents that watch, repair and escalate
GridPulse runs 20 stages every 15 minutes - ingesting, scheduling, settling and invoicing on its own schedule. Engines compute the numbers; agents decide what to raise and what can be safely repaired.
Every 15 minutes · 20 independent stages · Engines compute, agents decide
The Operating Cycle
Every 15 minutes the platform runs one pass of 20 independent stages. A failing stage is logged and the pass continues, so one bad feed cannot stop settlement or scheduling.
Invoices last
Rebuilt every pass, after the accounts they read are final.
Today is frozen
Once today's schedule exists it is never rewritten - deviation settles against it.
Weather history by age
Observed weather is pulled by the feed's own recorded age, not a fixed clock window.
Self-healing accounts
Any of the last three days short of 96 blocks is rebuilt automatically.
The Agent Model
An agent is either a monitor or an advisor - and that governs how its output is treated.
Monitors
Watch the platform continuously. What they report stays open until it clears, so it is tracked and appears in the operator's queue.
Advisors
Answer a question asked at that moment - a pre-gate briefing, or a what-if. Their answer describes a point in time, so it is never tracked as a standing defect.
How an agent works
Every agent is a deterministic procedure that may use a language model, never the reverse. One function gathers findings from the platform's own engines; another writes a plain summary without any model at all.
A model may reorder, group and explain those findings - but it cannot introduce a number. Any figure must already appear in the findings, or the computed wording is used instead. A model adds readability, not facts.
The Agents
Five monitors watch the platform's invariants every pass; two advisors answer on demand.
Operations Agent
Checks seven invariants each pass - schedule provenance, forecast coverage, settlement and metered completeness, feed freshness and cost agreement. It runs safe repairs and escalates when one keeps failing.
Trading & Scheduling Agent
Works against a fixed deadline: the day-ahead gate at 10:00 IST, where severity depends on time remaining. Checks gate readiness, price basis, deliverability, exchange limits, bid coherence, deviation exposure and the revision window.
Forecast Quality Agent
Measures signed forecast bias by day of week, then backtests a weekday correction on older data and scores it on newer data it never saw. Absolute error hides a systematic one.
Settlement & Dispute Agent
Ranks invoice discrepancies by rupees at stake, not percentage. Where a bill is worth disputing it prepares an evidence pack and draft letter - it sends nothing itself.
RPO & Carbon Compliance Agent
Measures the renewable purchase obligation against metered energy, excludes partial months and refuses to annualise. Prepares a regulator-ready statement for review before filing.
Trading Desk Analyst
Produces a pre-gate briefing: the day's position, the optimal split across the day-ahead, green day-ahead and real-time markets, the bid book with clearing confidence, and stress runs on the solver.
Scenario Copilot
Turns a question such as "what if coal costs rise 25%" into solver levers, runs it against live data and answers from the result. Where it cannot map the question, it says so.
Governed Autonomy
An agent can only re-run the platform's own engines over data it already has. It never changes a bid, invoice, dispute or rate, and it never sends anything.
A single switch governs unattended repairs. When off, every repair waits for approval. When on, only repairs marked unattended run without a person - anything touching money, the market or settlement always waits.
Complete audit trail
Every run is recorded with its findings, whether a model was used, and every action proposed or executed.
Attributed decisions
Approvals and dismissals record who decided, so the console can state whether a repair ran on a person's authority or its own.
Escalation on futility
A repair that runs without clearing its finding is not repeated indefinitely - after repeated attempts it escalates to a person.
No duplicate proposals
One open proposal exists per finding, so a persistent issue never accumulates identical requests.
Honest absence
Where the data cannot support a conclusion, the agent says so instead of producing a number.
See Feroe.ai in Action
Discover how AI agents can transform your operations and deliver measurable outcomes.