Vatavaran — Hirak Parekh

// THE PROBLEM

Every room, every summer afternoon.

You head out in the morning and the room is fine. By the time you get home, the afternoon sun has baked it. Your choices today are to run the AC all day for an empty room, or to walk into an oven and wait half an hour.

8 AM · YOU LEAVE

26°

The room is comfortable.

AFTERNOON · OUTSIDE HITS 35°

↑

The empty room heats up all afternoon.

5 PM · YOU'RE BACK

32°

Unless something saw it coming.

// HOW IT WORKS

Predict. Understand. Act early.

01 · PREDICT

Forecast the room

A deep-learning model learns how a room's temperature moves through the day and predicts where it's heading.

02 · UNDERSTAND

Talk to it like a person

"I'm cold" or "make it 75F" become precise temperature and fan changes. No menus, no remote.

03 · ACT EARLY

Cool it before you arrive

Using the forecast and the weather outside, it decides when to start cooling so the room is ready when you are.

// 01 · PREDICT

Learning a room's daily rhythm.

The model was trained on 1.5 million temperature readings: a year of minute-by-minute data from five sensor units.

90 engineered features: time of day and season as smooth cycles, recent readings, rolling averages and trends, and how fast things are changing.
A two-layer LSTM in TensorFlow, regularised so it learns patterns rather than memorising the year.
Handles sensors that disagree: the five units read up to 3.7°C apart, so each sensor's identity is part of the input and the model corrects for it.
Accurate to about 0.3°C on data it never saw during training.

Line chart of predicted indoor temperature over a day, rising from about 21°C in the morning to a 29.9°C peak in the afternoon and falling in the evening — A full day, forecast: the room climbs from about 21°C in the morning to almost 30°C by early afternoon. That curve is what tells the system when to act.

Command parser window: the input 'set temperature to 30 degrees' becomes a structured result with an absolute temperature of 30 and confidence 0.98 — The command parser turning plain English into a structured instruction the AC can act on, with a confidence score.

// 02 · UNDERSTAND

"I'm cold" is a complete instruction.

People don't think in setpoints. Vatavaran reads how you feel and how strongly, and turns it into an exact change, adjusting the fan along with the temperature.

"I am cold"→+2°C · fan down one

"very hot"→−3°C · fan up one

"make it 75F"→set to 23.9°C

"set fan to high"→fan level 3

// 03 · ACT EARLY

Start cooling at exactly the right time.

The control logic compares where the room is heading with where you want it, checks the live weather forecast, and works out how early to start. Hot days get a head start; mild days don't waste power.

HOT DAY · GAP OVER 10°C

2 hours early

Starts cooling well before you arrive, so the room is ready on time.

WARM DAY · GAP 5–10°C

90 minutes early

A shorter head start for a smaller gap.

MILD DAY · GAP 2–5°C

30 minutes early

Just enough, with no hours of cooling an empty room.

And when you override it, it remembers to undo it: after a few hours it eases back to the normal schedule, so a forgotten "make it 18" doesn't run all night.

// RESULTS

What it adds up to.

±0.3°C

Prediction error on data the model never saw in training

44%

Less AC energy than running it all day, in a simulated day of smart scheduling

1.5M

Real sensor readings the model learned from, across a full year

<100ms

To understand a spoken-style command and turn it into an action