CoolSense: an AC that tells you what's wrong, before the technician drives out

Malathi Venkatesan — Sun, 18 Oct 2026 17:43:18 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

What I Built

The AC stops cooling. A technician finds a gas leak and says: "I'll come back tomorrow with the parts." One fault, two visits.

CoolSense lets the AC say what's wrong first:

① Detect (code): spots when this unit drifts from its own normal.
② Diagnose (TabPFN): names the fault.
③ Explain (Gemma, local): messages the owner in their language. The technician gets what to bring.

Who pays: AC maintenance companies, per AC per month. They see which unit needs a visit, what's wrong, and what to put in the van.

Demo

Real sensor data from a rooftop AC with a documented 40% gas leak (Berkeley Lab), replayed as if live. Runs on my laptop.

Code

malathilatha / coolsense

AC fault detection that runs on a laptop: code spots the drift, TabPFN names the fault, and local Gemma explains it to the owner in their language, tested on real LBNL rooftop-AC data with a documented gas leak

❄️ CoolSense: the AC that tells you what's wrong

Code detects → TabPFN diagnoses → Gemma explains. All open models, all on one laptop, no cloud.

When an AC fails, a technician usually visits once to find the fault and again with the parts CoolSense watches the unit's own sensor readings, names the likely fault, writes the owner a plain message in their language, and gives the technician a job card with what to bring, so the first visit can fix it.

Who pays: AC maintenance companies and facility managers (per AC, per month). They get a priority list of which units need a visit, what's probably wrong, and what to put in the van.

Built for the DEV Hacktoberfest Open-Source AI Challenge, Week 2: Side Hustle.

Demo

Live AC (owner + technician)	Maintenance company (B2B)

The app replays real recordings from a rooftop AC with a documented…

View on GitHub

How I Built It

Detect: gradient boosting learns what each sensor should read for the weather and load. Big gaps raise a flag.
Diagnose: TabPFN classifies flagged readings from labelled examples. Trained on simulation only, it got the real leak wrong. Add 2 technician-confirmed days, with no retraining, and it named 98% of flagged readings correctly.
Explain: gemma3:4b via Ollama writes the message. It never picks the fault or the numbers. The job card comes from a fixed checklist.

Real AC	Days flagged
Leak, June 2020	11 / 15
Healthy circuit, 2021	2 / 109
"Healthy" circuit B, 2021	53 / 53: looks like the leak came back (hypothesis)

Not yet: it's a replay. Next comes reading live data from building systems. Newer TabPFN weights are non-commercial, so the code has a gradient-boosting fallback.

Why Does Open Innovation Matter?

Open data (LBNL, CC BY 4.0) let me test on a real leak.
Open weights run beside the AC: no cloud bill per unit, no building data leaving the site.
A technician's confirmation updates TabPFN on the spot, with no vendor in the loop.

Prize Categories

TabPFN: diagnoses the fault and improves from 2 confirmed days.
Gemma: writes the owner's message locally in 7 languages.

PathShala AI: Offline, Zero-Install Teaching Assistant

Malathi Venkatesan — Sun, 01 Mar 2026 18:25:37 +0000

This is a submission for the DEV Weekend Challenge: Community

The Community

I grew up watching teachers struggle in classrooms where the internet is either absent or unreliable. In rural Area, millions of school teachers face the same problem every single day: they are expected to deliver quality, engaging lessons — but the digital tools that could help them are locked behind a stable internet connection they simply don't have.
Free AI tools like ChatGPT and Gemini exist — but they're completely useless when there's no internet. Offline AI tools like Ollama and LM Studio exist — but they require terminal commands, admin rights, and software installation that a government school teacher cannot do.

What I Built

PathShala AI is a browser-native, zero-install AI teaching assistant that runs 100% offline after a one-time setup — no app, no admin rights, no internet, no API key. Ever.
A teacher downloads one HTML file. Opens it in Chrome. Done. From that point, the full power of a local LLM is available — in a classroom with no WiFi, in a village with no mobile data, even on a flight.

Why a Single HTML File?
Rural school computers run older Windows with no package manager, no Node.js, and no IT support. A single HTML file that a teacher can open in Chrome is the only deployment format that works in this reality.
Installation is: download file → double-click. That's it.

Demo

Watch Demo Video

Code

PathShala AI – HTML Source

How I Built It

Tech Stack:
WebLLM (Apache TVM) – In-browser inference
Llama 3.2 1B (q4f32_1-MLC) – Quantized model
WebGPU – Hardware acceleration
Cache API + IndexedDB – Full offline storage (~900MB)
Web Speech API – Offline TTS
Vanilla HTML/CSS/JS – Zero dependencies

Architecture
Query → Local Model (Cache) → WebGPU Inference → Text/Voice Output
No backend. No API key. No data leaves the device.

Team Submissions: Build By Malathi Venkatesan
username: malathilatha199615

DEV Community: Malathi Venkatesan

CoolSense: an AC that tells you what's wrong, before the technician drives out

What I Built

Demo

Code

malathilatha / coolsense

AC fault detection that runs on a laptop: code spots the drift, TabPFN names the fault, and local Gemma explains it to the owner in their language, tested on real LBNL rooftop-AC data with a documented gas leak

❄️ CoolSense: the AC that tells you what's wrong

Demo

How I Built It

Why Does Open Innovation Matter?

Prize Categories

PathShala AI: Offline, Zero-Install Teaching Assistant

The Community

What I Built

Demo

Code

How I Built It