Intelligence

Six specialists. One argument, settled by data.

Gen318 isn't a dashboard with AI bolted on. Below is an actual decision — six agents reasoning together over a Harmattan front, 72 hours before it hits your cluster.

Agent coordination · LangGraphReasoning

Decision trace · Adamawa clusterLIVE · 06:02 WAT

WEATHER06:02:11

Harmattan front reaching Adamawa cluster in ~72h. PV yield −35 to −48%.

DEMAND06:02:14

Thursday is Ganye market day. Load +22% against a reduced-solar window.

DISPATCH06:02:19

Re-planning 72h horizon: deep-charge tonight on full sun, defer genset to Thu 17:40–20:10.

BATTERY06:02:21

Constraint: string B2 max depth 78% — cell divergence watch. Accepted.

DISPATCHDECISION06:02:26

Plan committed inside guardrails. Projected diesel saved: ₦38,400. Cleaning crew suggested for Fri.

COMPLIANCE06:02:27

Logged to audit trail. Q2 uptime projection unchanged: 96.2%.

Specialist agents

<50ms

Edge inference

71→86%

Accuracy, 6 months

₦38,400

One decision, one site

01 · The agents

Each one reads different instruments.

A battery chemist, an inverter electrician, a market economist, a dispatcher, an auditor, and a meteorologist — as models. Each card shows what its agent reads, and the call only it can make.

Battery Health

LSTM · XGBOOST EDGE

Reads: Charge cycles, impedance, cell temps
Calls: Degradation vs expected · remaining useful life

West-Africa heat-cycling model v3

Inverter Diagnostics

WAVEFORM ML · 12 FAULT CLASSES

Reads: Voltage / current signatures, efficiency curves
Calls: Failure windows 2–4 weeks out

IGBT drift flagged at Adamawa-07

Demand Forecast

PROPHET + LSTM

Reads: Load history, weather, market calendars
Calls: Hourly → monthly load, with confidence bands

Knows market days, Ramadan, harvest

Dispatch Optimization

RL · STABLE BASELINES3

Reads: Forecasts, tariffs, fuel price, battery state
Calls: The cheapest safe solar–battery–diesel mix

15–30% more value than rule-based

Compliance

GENERATIVE · TEMPLATED

Reads: Uptime, power quality, connections, milestones
Calls: Odyssey-ready reports · holdback risk

2–5 days of work → under an hour

Weather

NASA POWER + LOCAL MET

Reads: Irradiance, dust, rain, heat forecasts
Calls: Harmattan soiling, flood risk, heat stress

Pre-positions crews before storms

02 · Compound learning

Accuracy is a metric we report, not a claim we make.

Every closed work order is a labeled training example — what was predicted, what the technician actually found, what fixed it. Your fleet's accuracy curve ships in your quarterly report, next to uptime.

And because twins pre-train on synthetic failures, site 50 starts smarter than site 1 ever was — day-one accuracy of 70–80% instead of a cold start.

Prediction accuracy · fleet-wide · reportable metricMONTHS 0–6

87%

inverter-failure accuracy @ 14-day horizon, 150 sites

70–80%

day-one accuracy for new sites, from fleet priors

8–12 pts

industry benchmark gain, month 3 → 12 (OxMaint)

03 · Where the models run

The edge acts. The cloud learns.

Heavy models train on fleet-wide history in the cloud, then distill down to XGBoost that runs on the site's own hardware. When the network disappears — and it will — the intelligence doesn't.

At the site · edge nodeACTS

Distilled XGBoost models
anomaly + dispatch inference in <50ms
Local dispatch authority
keeps optimizing with zero connectivity
30+ day buffer
store-and-forward, 4-tier priority sync
Runs on 2G
or no G — rural connectivity is the design case

In the cloud · fleet brainLEARNS

LSTM · Prophet · RL training
heavy models learn on fleet-wide history
A digital twin per site
simulates futures, generates synthetic failures
Federated learning
cross-operator gains, no raw data shared
Distill → deploy
cloud learning ships back to every edge node

04 · Conversational operations

Ask the fleet why. It shows its work.

The same orchestrator that coordinates the agents answers your team — over WhatsApp, SMS, or the console. Not canned responses: it decomposes the question, queries the agents, and returns the answer with the trace behind it.

Operators ask why — and get causes ranked by impact
Actions by reply: work orders, schedules, approvals
Investors ask for portfolio numbers in plain language
Works where your team already is — no new app to learn

Gen318 Ops

WhatsApp · your fleet's number

Online

Why did the genset run 6 hours at Adamawa-07 yesterday?

07:14

Three causes, in order of impact:

· Cloud cover cut PV yield 38% vs forecast
· Market day pushed evening load +22%
· I preserved battery at 24% — string B2 is on a divergence watch

Net cost vs doing nothing: ₦41,200 saved. Full trace available.

07:14 · answered in 4s

Schedule the panel cleaning before Friday.

07:15

Done — WO-0147 assigned to Musa I. (94% first-visit rate), Thursday 09:00.

WORK ORDER CREATED

05 · The AI-native stack

Seven systems that compound.

Each one feeds the others — twins train the predictors, field visits label the data, federated learning spreads the gains. That loop is the moat: it can't be copied without the fleet.

02 · Autonomous digital twins

Every night, each site lives a thousand tomorrows.

The twin doesn't just mirror the site — it imagines its futures. It simulates dispatch policies before they run, and generates the synthetic failures that let predictive models train before a single real one occurs. That's how new sites skip the 12-month cold start.

Digital twin · Kaduna-12 · tonight's simulation run1,000+ FUTURES

12 mo → weeks

time-to-accuracy, via synthetic failure pre-training

Recalibrated

every telemetry cycle — design vs as-operated

+10kW PV?

capex scenarios answered from the living model

Cross-site coordination · MARL · interconnected clusterNegotiating

Without coordination, three sites run three gensets. With it, one runs and exports — each agent optimizes locally, negotiates globally.

03 · Multi-agent RL coordination

Interconnected sites stop acting alone.

As DARES clusters interconnect, optimization stops being a single-site problem. Each site's agent makes its own decisions — and negotiates with its neighbours over exports, shared genset duty, and mesh load balancing.

05 · Self-healing operations

The 2 AM failure your team reads about at 7.

Self-healing operations · incident WO-0151 · while you sleptRESPONSE: 7s