Blog

Notes from building it

Benchmarks, the parts of Ge'ez typography that break every default, and the billing decisions that are easier to explain in prose than in a pricing table.

engineering/2 October 2026/5 min

Four promises behind one endpoint

Metered inference is easy to get wrong in ways a customer only finds on their invoice. Here are the four rules our request handler follows, and why each one is there.

24 September 2026
16.1% word error, and what that number hides
A benchmark number is a claim about a test set, not about the world. Here is the test set, and here is where our model is still much worse.
research / 8 min
12 September 2026
Caption timing for a script that is not Latin
Every subtitle tool assumes characters per second measured on English. Ge'ez packs a consonant and a vowel into one glyph, and the usual numbers are wrong by about half.
engineering / 7 min
28 August 2026
Credits, not subscriptions
A recurring card is a payment method half our market does not have. Everything downstream of that assumption breaks.
product / 5 min
9 August 2026
A model small enough for the phone in the room
Our ASR runs in a browser tab with no GPU. That constraint shaped the model more than any accuracy target did.
engineering / 6 min

Start with the one that is already in production

Amharic transcription, two free hours a month, one endpoint and no SDK to install.

Read the quickstart Talk to us

Four promises behind one endpoint

16.1% word error, and what that number hides

Caption timing for a script that is not Latin

Credits, not subscriptions

A model small enough for the phone in the room

Start with the one that is already in production