Ampera — AI-Powered Predictive Maintenance for EV Charging Infrastructure

Inspiration

The idea for Ampera came from a frustrating truth about EV charging: failures are often silent until they become public. A charger can degrade for hours (or days) — voltage drifting, temperature creeping, error codes ticking up — and no one notices until a driver shows up, gets an out-of-service screen, and leaves angry.

That moment is bigger than a bad UX. Every broken charger is a tiny vote against EV adoption.

So we asked a simple question:

Why are we waiting for outages to tell us something is wrong, when the signals are already there?

Ampera is our answer: turn reactive maintenance into predictive maintenance, using live telemetry + anomaly detection + an AI assistant that explains what’s happening and what to do next.

The Problem

EV charging networks are critical infrastructure, but they fail constantly and silently.

Operators often learn about downtime from customer complaints, not their own monitoring.
Reactive workflows mean expensive truck rolls, slow triage, and repeat failures.
Each hour of downtime costs revenue, erodes trust, and discourages EV adoption.

Traditional monitoring is mostly thresholds and alerts — after something breaks.

Ampera makes it predictive.

The Solution

Ampera is a real-time intelligence platform that:

Monitors live charger telemetry (voltage, current, temperature, session duration, error codes)
Detects patterns indicating early failure using a machine learning model
Produces a continuously updated risk score per charger
Automatically logs anomalies and incidents for auditability
Provides an AI triage assistant that explains why and recommends exact actions

The goal is simple:

Catch the failure before the driver ever sees it.

Core Features

1) Live Network Dashboard

A fleet-wide, at-a-glance view of charger health:

Green = healthy
Yellow = watch
Red = act now

Operators don’t dig through logs; they see risk instantly.

2) Predictive Anomaly Detection

Ampera continuously scores each charger for failure risk using unsupervised anomaly detection:

Input: streaming telemetry (with realistic noise)
Model: Isolation Forest (scikit-learn)
Output: ( 0 \rightarrow 100 ) risk score, refreshed every few seconds

Instead of “alert when it’s broken,” we aim for:

predictive signal → maintenance action → no downtime

3) AI Triage Assistant

An embedded chat assistant powered by an LLM (Claude or OpenAI API).

Operators ask:

“Why is charger 12 flagged?”
“What should I do about the overheating unit in Lot B?”
“Is this likely a sensor issue or a real thermal problem?”

Each query is injected with the charger’s recent metrics + anomaly type, so responses are specific, actionable, and not generic.

4) Incident Timeline

Every anomaly, escalation, and resolution is logged with:

timestamps
severity
who acted and what they did

This creates an automatic audit trail without extra paperwork.

5) Charger Detail View

Click any charger to see:

live metric graphs (voltage / temperature / sessions / errors)
current risk score
model “reasoning” signals (e.g., which metrics drifted)

How We Built It

System Architecture (Demo-Ready)

Everything runs locally for hackathon reliability and speed.

Simulator generates real-time telemetry for 20 chargers
ML pipeline calculates anomalies + risk scores
FastAPI backend serves REST endpoints to the dashboard and logs incidents
React ops dashboard renders fleet health, detail graphs, and timelines
LLM assistant answers questions using charger-specific context injection

Tech Stack

Frontend

React + Tailwind CSS
Recharts for live charts
Dark-mode ops dashboard aesthetic (fast scanning, low noise)

Backend

Python + FastAPI
REST endpoints for:
- telemetry stream
- risk scores
- incident logs
- assistant query endpoint

Data Simulation

Python script streaming telemetry for 20 chargers
3–4 chargers have injected failure patterns:
- gradual voltage drops
- temperature spikes
- increasing error rates

Machine Learning

Isolation Forest for unsupervised anomaly detection
Risk scoring pipeline converts model output into a human-readable health score

AI Assistant

Claude or OpenAI API
Prompt injection pattern:
- recent metrics window (e.g., last 10–15 minutes)
- current risk score + direction (rising/falling)
- detected anomaly category (thermal / voltage drift / error burst)
- operational playbook-style response format

Risk Scoring (Concept)

Isolation Forest produces an anomaly score based on how “different” the current telemetry looks compared to typical behavior.

We then map that into a (0!-!100) risk score:

$$ \text{risk} = \text{clip}_{0}^{100}\Big(\alpha \cdot \text{anomaly_score} + \beta \cdot \text{trend_penalty}\Big) $$

anomaly_score: how unusual the current point/window is
trend_penalty: extra weight if key metrics are drifting steadily (the “silent failure” signature)
clip keeps it bounded and dashboard-friendly

Finally, we bucket for ops clarity:

0–39: Healthy (Green)
40–69: Watch (Yellow)
70–100: Act (Red)

Challenges We Faced

1) Making simulated data feel real

Random noise isn’t enough — real infrastructure fails with patterns. The hardest part was designing failure injection that looked believable:

slow drift (voltage sag over time)
correlated changes (temperature + error rate rising together)
intermittent resets (fake “it went away” moments that fool threshold alerts)

2) Avoiding “anomaly spam”

Unsupervised models can over-flag. We had to tune:

contamination / sensitivity
windowing strategy
smoothing of risk scores so the dashboard doesn’t flicker

We wanted “predictive and calm,” not “noisy and anxious.”

3) Getting the assistant to be actionable

LLMs love being verbose. Operators need:

the likely cause
severity
next actions (what to check, what to reboot, what to dispatch)
when to escalate

Prompt structure mattered more than model choice.

4) Real-time UX

A live ops dashboard can become chaotic quickly. We focused on:

stable layout
clear color semantics
fast drill-down
incident log that reads like a story, not a dump

What We Learned

Predictive maintenance is a product problem, not just an ML problem.
The model is useless if the operator can’t act confidently.
Explainability beats raw accuracy in demos.
Judges (and operators) trust systems that show why.
Time-series “drift” is where the value is.
A single spike is easy. Catching “it’s slowly getting worse” is the win.
LLMs shine as a bridge between signals and actions.
Telemetry → plain-English triage → checklist actions is the magic.

Team & Division of Work

Backend Dev 1 — Data & ML

Built telemetry simulator (20 chargers, noise + injected failures)
Trained/tuned Isolation Forest
Built risk scoring pipeline
Output clean JSON for API consumption

Backend Dev 2 — API & AI

Built FastAPI endpoints for telemetry, scores, and incidents
Integrated LLM triage assistant
Wrote system prompt + context injection format
Incident creation logic when thresholds are crossed

UI/UX Designer — Frontend & Presentation

React dashboard (network grid, detail view, charts)
Chat assistant sidebar
Incident timeline feed
Devpost page, screenshots, and demo narrative

Build Timeline (22 Hours)

Hours 0–2: Setup, simulator emitting data, API skeleton, wireframes
Hours 2–6: Model trained, endpoints live, dashboard shell consuming API
Hours 6–12: Live scoring + color coding, detail graphs, assistant working
Hours 12–18: Incident timeline, refined assistant, UI polish + edge cases
Hours 18–22: Full demo rehearsal, Devpost + README, speaker flow locked

Demo Narrative (2 Minutes)

“340 million EVs are projected to be on the road by 2030. The infrastructure holding them up is breaking silently every day. Ampera fixes that.”

Open the dashboard: show a mostly green network.
Watch charger 7 trend yellow as voltage drops and temperature rises.
Ampera flags it and creates an incident automatically.
Ask the assistant:
“What’s wrong with charger 7 and what should I do?”
Show the response (cause + steps) and the metric history proving the trend started hours ago.
Close:
> “Ampera caught this before any driver ever saw an out-of-service screen.”

Track Strategy

Best Data Science: emphasize anomaly detection + time-series drift + risk scoring
Best Design (With Code): the dashboard is the differentiator — instant clarity
Best Social Good: uptime is EV adoption; this is climate infrastructure
Best Overall: full-stack + ML + LLM + polished UI in 22 hours