ElevatorInsight

Inspiration

The inspiration for ElevatorInsight came from a simple observation: elevators are part of our everyday lives, yet we usually only think about them when something goes wrong. A breakdown can cause inconvenience at best, and safety risks at worst. During our hackathon brainstorming, we asked ourselves:

What if we could catch early warning signs of malfunction before they turn into real problems?

That question set the stage for building a system that transforms raw IoT sensor data into actionable, real-time insights.

What it does

ElevatorInsight is a real-time monitoring system that takes in raw elevator sensor data, evaluates the risk of malfunction using a machine learning model, and translates those results into clear, actionable alerts. Instead of just flagging anomalies with numbers, it explains the why behind them by using SHAP values and a language model to summarize the issue in plain English. These alerts are then instantly delivered to Slack for operators and mirrored on a lightweight web dashboard, helping teams spot early warning signs, reduce downtime, and keep elevators running safely and reliably.

How I built it

Data Replay: I used a replay script to stream rows from a dataset of engineered features as if they were live IoT signals.

Risk Scoring: Each row was passed to an ML model via a FastAPI service, producing a risk probability.

Explainability: I extracted SHAP features to highlight what drove the model’s decision.

Summarization: A LangChain agent turned raw numbers into a single human-readable sentence (e.g., “monitor brakes.”).

Notifications: The alerts were pushed both to Slack and mirrored to a lightweight web dashboard built with HTML + Tailwind CSS.

Challenges I ran into

Data Volume: Real datasets are huge. Iterating without waiting hours meant building a sparse replay mode to skip straight to anomalies.

Static Hosting: GitHub Pages doesn’t serve live file changes. We had to implement an auto-commit mechanism to push updates for the web dashboard.

Integration Puzzles: Getting different components (Python scripts, JSON writes, JS polling) to talk to each other smoothly was trickier than expected.

Humanization of Alerts: Translating model output into something a non-technical operator can trust required balancing detail with clarity.

Accomplishments that I'm proud of

One of my biggest accomplishments is successfully stitching together a complete end-to-end pipeline in just a short hackathon timeframe. I'm proud that I didn’t just stop at building a model—I went further to make it explainable, human-readable, and accessible through Slack and a live dashboard. Getting SHAP explanations to flow into a LangChain agent and then pushing those insights out as natural language alerts felt like a real breakthrough. I also take pride in solving practical hurdles—like replaying only anomalous rows to save time, or auto-committing updates so the web page would refresh—which showed me I could balance technical depth with usability. Ultimately, I'm proud that ElevatorInsight went from an idea on paper to a working system that could genuinely improve elevator safety and reliability.

What we learned

How to design a real-time pipeline that continuously ingests, processes, and interprets data.

The importance of explainable AI—it’s not enough for a model to say “something’s wrong,” it needs to tell operators why.

How to combine multiple tools (ML models, SHAP values, and LangChain agents) into one cohesive flow.

The quirks of working with static sites (GitHub Pages) and ensuring live updates are visible to end users.

What's next for ElevatorInsight

Next for ElevatorInsight is turning my prototype into a scalable, production-ready platform. I plan to integrate real-time data streams using technologies like Kafka or WebSockets instead of simulated CSV replay, so alerts truly happen the moment an anomaly is detected. I also want to expand the dashboard from a single alert view into a full monitoring console with historical trends, interactive charts, and multi-elevator comparisons. On the AI side, I'll explore predictive maintenance models that don’t just flag current anomalies but forecast potential failures days in advance. Finally, I envision extending the system beyond elevators to other critical IoT infrastructure—like escalators, HVAC systems, or even factory equipment—making ElevatorInsight a general-purpose safety and reliability companion.

Built With

• aiohttp-machine-learning:-custom-risk-scoring-model-with-explainability-via-shap-values-apis-&-integrations:-slack-webhooks-api-for-real-time-alerts-platforms-&-tools:-github-pages-(for-dashboard-hosting)
• css
• css-(tailwind-css)-frameworks-&-libraries:-fastapi
• fastapi
• frontend
• git/github-(version-control-&-auto-deploy)-data-handling:-pandas-(for-replaying-and-filtering-csv-sensor-data)
• html
• javascript
• json-for-alert-storage-&-exchange-other:-async-pipelines-with-python-asyncio
• langchain
• lightweight
• polling
• python
• shap