DEV Community

Cover image for BatteryIQ
Anushka Singh
Anushka Singh

Posted on

BatteryIQ

Our idea for the Tek Qubit hackathon by Tekion, and how we plan to build it in public.

Why we're writing this

We're MLOGs, a team of 3 final year B.Tech students , and we're taking part in Tek Qubit, a hackathon by Tekion, a cloud platform company for automotive retail. This post covers our idea and our plan. We'll update it as we build, including the parts that go wrong.

The problem we picked

When a dealership takes a used EV as a trade-in, the battery is the most important part of its value, and also the hardest to judge. An appraiser often has no clear picture of its health. That creates three problems:

  1. The dealer overpays for battery life that isn't there.
  2. The dealer underpays and loses the customer.
  3. The customer can't tell how the number was reached.

We chose this from the Sustainability theme because a fair valuation also helps EVs stay in use longer instead of ending up as waste.

Our idea: BatteryIQ

BatteryIQ estimates an EV battery's state of health (SOH) from data like charge cycles, mileage, age, temperature history and service records. The output is a report, not just a number:

  • SOH % and estimated remaining life
  • A confidence score, so people know how much to trust the estimate
  • The key reasons behind it (using SHAP feature explanations)
  • A fair value range for the trade-in

If confidence is low, the tool says so and recommends a manual inspection. We think a tool that admits uncertainty is more useful to an appraiser than one that always sounds sure.

How we plan to build it

  • Language and libraries: Python, XGBoost, SHAP, Streamlit
  • Data: public battery-aging datasets plus synthetic service-history data
  • First step: a simple baseline model. Our XGBoost model has to beat it, or we'll say so.
  • Success metric: lower prediction error than the baseline on held-out data ## The limits, stated up front

Most public battery datasets come from lab tests on cells, not from real vehicles on the road. So our prototype can show that the approach works, but it can't prove accuracy on real fleets. A real deployment would need diagnostic data from dealers or manufacturers. We'd rather say this now than have it hidden in the demo.

Where it fits

We imagine BatteryIQ as a step inside a dealership's existing appraisal workflow, so an appraiser sees the battery report where they already work, instead of opening a separate app.

What's next

  • [ ] Choose and clean the datasets
  • [ ] Build the baseline
  • [ ] Train the model and add the confidence score
  • [ ] Add SHAP reasons and the value range
  • [ ] Build the demo

We'll share what we learn in the next post, including our dataset choices and our first baseline results.

Top comments (0)