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.
Harmattan front reaching Adamawa cluster in ~72h. PV yield −35 to −48%.
Thursday is Ganye market day. Load +22% against a reduced-solar window.
Re-planning 72h horizon: deep-charge tonight on full sun, defer genset to Thu 17:40–20:10.
Constraint: string B2 max depth 78% — cell divergence watch. Accepted.
Plan committed inside guardrails. Projected diesel saved: ₦38,400. Cleaning crew suggested for Fri.
Logged to audit trail. Q2 uptime projection unchanged: 96.2%.
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
Measured against the approved site baseline
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
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.
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.
- Distilled XGBoost modelsanomaly + dispatch inference in <50ms
- Local dispatch authoritykeeps optimizing with zero connectivity
- 30+ day bufferstore-and-forward, 4-tier priority sync
- Runs on 2Gor no G — rural connectivity is the design case
- LSTM · Prophet · RL trainingheavy models learn on fleet-wide history
- A digital twin per sitesimulates futures, generates synthetic failures
- Federated learningcross-operator gains, no raw data shared
- Distill → deploycloud learning ships back to every edge node
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
Why did the genset run 6 hours at Adamawa-07 yesterday?
07:14Three 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 4sSchedule the panel cleaning before Friday.
07:15Done — WO-0147 assigned to Musa I. (94% first-visit rate), Thursday 09:00.
WORK ORDER CREATEDSeven 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.
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.
Without coordination, three sites run three gensets. With it, one runs and exports — each agent optimizes locally, negotiates globally.
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.
The 2 AM failure your team reads about at 7.
Kaduna-12 inverter offline. Related signals grouped into one critical incident.
Safety interlocks pass. Enabled sites start backup automatically; advisory sites route the same action for approval.
Amina T. ranked by skill, distance, availability, and workload. Work order assigned; IGBT module suggested.
SMS and push escalation sent immediately. Routine follow-up is held for the morning brief.
Reviews command readback, incident evidence, technician assignment, and the remaining follow-up.
One evidence trail closes the loop from detection to verified action. Each site graduates from monitor to advisory, supervised, and full automation under explicit guardrails.
Generative narratives
Weekly operations analyses written by the system — findings, causes, and what to do next, not chart dumps.
Federated learning
Models improve across operators without pooling anyone's raw data. Everyone's fleet gets smarter.
Computer-vision inspection
Thermal + visual imagery classified panel-by-panel — hot spots, micro-cracks, soiling — fed straight into the twin.
01 conversational ops — demonstrated above · 02, 03, 05 — shown in full
Every insight arrives with a price tag.
The agents don't just detect — they rank what they find by naira at stake, profile every site against the fleet, and show you where the next intervention pays best.
Genset staging across Adamawa cluster is mistimed vs market days — reschedule recovers 9% of diesel spend.
DISPATCHTwo strings trending to replacement in the same quarter — stagger now to smooth capex.
BATTERY14 customers show pre-default top-up patterns at Niger-03 — engage before disconnection.
REVENUEPost-Harmattan cleaning is 11 days late on average — soiling losses compounding.
WEATHERDiesel share of generation · trailing 8 weeks
Watch the agents argue over your own sites.
Free fleet assessment — anomalies surfaced from your existing VRM data.