Chai & Kettle

Reverie · Low-data NLU for voice actions

Shipped a deep-learning NLU engine for voice actions to over 3 million devices, reaching about 99% accuracy from only a handful of examples.

01 — Project context

What exists today, and how this does it better.

FIG. 09

Overview

A leading set-top-box provider wanted to run actions from voice. The speech team handled the ASR; we built the part that decides what you actually meant.

At Reverie, a former colleague — and dear friend — proposed we build the NLU engine together. We did, in marathon mode: the model, and the whole platform around it.

01 — The split

From voice to action

Speech becomes text (ASR), text becomes intent (NLU), intent becomes an action on the device. The speech team handled ASR — on-device and cloud. We built the NLU layer — the part that turns 'put on the cricket' into a command the box can run.

02 — The model

Intent from almost no data

The hard constraint: recognise intent from a handful of examples, in any language. This was before transformers reshaped NLU — so the toolkit was multilingual word and sentence embeddings (LASER's BiLSTM encoder, fastText subwords, MUSE's aligned vector spaces) over a BiLSTM / CNN classifier, with BPE keeping it language-agnostic. We built a deep-learning engine on those ideas — handling multi-intent, with auto-annotation to stretch every example further.

LASER
fastText
MUSE
BPE
BiLSTM
CNN

03 — The platform

Create a class, click train, ship

Chai was the backend brain; Kettle the platform on top. Create a project, add intent classes, add a few examples, click to train, deploy, and test the new model — a no-code loop a non-ML team could run themselves. Teach-me / test-me, auto-annotate, repeat.

04 — Operations

Zero-downtime by design

New models shipped constantly, so we built the AutoML ops to swap them in with zero downtime — traffic never dropped while a freshly trained model took over. Retraining stopped being an event and became routine.

Outcome

~99%
Accuracy across all set-top-box use cases
3M+
Devices across India

Almost no data in. Ninety-nine percent out — across three million devices.

My role

Built the NLU engine and the Chai & Kettle platform with a longtime collaborator, model, training loop, and the AutoML ops that let new models ship with zero downtime. Delivered well past target and into millions of living rooms.

Outcome

Content coming soon.

Stack

Deep NLU
BPE
Multilingual
AutoML Ops
Full-stack