All projects
Renewable Energy · Agentic AI + Optimization

AEOLUS — Renewable Fleet Operations Brain

A governance-first multi-agent system that runs a wind fleet like a great operations director — detect, diagnose, price against the live market, schedule with a real solver, and gate every action through human approval.

The problem

Everyone builds agents that predict turbine failures. Prediction is commoditised. The hard, valuable problem is the closed loop after the prediction — deciding when to act so you lose the least generation revenue, under real constraints (crew, weather, grid commitments).

And a regulated operator cannot deploy an agent that just acts. It needs economically-optimal scheduling against a live electricity market and an immutable governance trail it can actually sign off — a policy gate, a simulation pre-check, human approval, and an auditable record of every decision.

Architecture

  1. 1

    Data plane

    Real Kelmarsh SCADA telemetry + German day-ahead market (energy-charts) + Open-Meteo hub-height weather.

  2. 2

    Lakehouse

    Bronze → Silver (+ ISA-95 asset registry) → Gold feature store — a parquet medallion.

  3. 3

    Perception

    Normality models, anomaly + prognosis (health score + lead-time), power-curve/generation forecast, SHAP-style attribution — trained on the clean baseline only.

  4. 4

    Cognition (LangGraph mesh)

    Orchestrator → Diagnostician → Market → Scheduler/Optimizer → Work-order, with RAG over the O&M manuals.

  5. 5

    Optimization core (OR-Tools CP-SAT)

    Picks the maintenance window t* minimising LostRevenue(t) + RiskCost(t) under crew, weather (≤12 m/s safe-climb), and grid-commitment constraints; incidents compete for shared crews via optional-interval no-overlap.

  6. 6

    Governance & action

    OPA-style policy gate · digital-twin sim pre-check · one-click human approval · immutable hash-chained audit log · fleet-wide kill switch.

Key tradeoffs

The LLM reasons and explains; OR-Tools does the optimisation.

Why · Never ask a language model to do the math it is bad at. The Scheduler agent reads the solver's answer and narrates the rationale — it does not invent the schedule.

A CP-SAT model with optional-interval no-overlap, not a hand-rolled argmin.

Why · Multiple incidents genuinely compete for shared crews; only a constraint solver handles that combinatorics correctly, and a naïve "fix-now" baseline is solved too for an honest delta.

Report value as two honest levers — generation revenue protected + unplanned-failure cost avoided.

Why · The build surfaced that the safe-climb constraint excludes the windiest (highest-generation) windows, so the lost-revenue spread among climbable windows is modest; the dominant value is catching the fault early for planned-vs-run-to-failure. Showing both is the honest framing.

Governance as a first-class layer, not a bolt-on.

Why · A regulated operator can only deploy autonomy whose every action is gated, simulated, human-approved, and auditable — so the policy gate, hash-chained log, and kill switch are core, not afterthoughts.

Eval results

two honest levers
Value protected

Generation revenue saved by the cheapest safe OR-Tools window + P(failure)×(unplanned−planned), measured against a solved naïve fix-now baseline.

real + free
Data

Kelmarsh SCADA (6× Senvion MM92, 2016, CC-BY-4.0) scaled to a 20-turbine fleet; live DE day-ahead prices + Open-Meteo. $0 to run.

genuine
Learned signals

Normality models trained on the clean baseline only, so residuals, health score, prognosis lead-time, and attribution are genuinely learned — not read off the injected scenario.

CP-SAT
Optimizer

Google OR-Tools constraint-programming scheduler with crew/weather/grid constraints; multi-incident crew contention via no-overlap.

Production proof

The artifact that keeps the numbers honest — the eval harness / monitoring gates that run in CI, not a one-off notebook result.

Governance trail + digital-twin gate

CI · passing
Policy gate (OPA-style) checked
Digital-twin sim pre-checked
Hash-chained audit log live-verified
Fleet-wide kill switch armed

Every agent action passes a policy gate and a digital-twin simulation, then waits for one-click human approval; the reasoning log is hash-chained and verifiable, with a fleet-wide kill switch — the controls a regulated operator needs before deploying autonomy.

Tackles the part of industrial AI that actually pays — not predicting failures (commoditised) but the economically-optimal, governance-gated closed loop that decides when to act — on real European wind data, for $0.

Let's talk

I'm focused on finance AI — credit risk, RegTech, AML, and agentic investment research. Open to roles, mentorship, and collaborators in fintech, quant, and bank AI.