AI coding tools have completely changed how fast we ship.
For a small team, they are an insane force multiplier.
Writing / Essays / Notes
Notes from building voice and language infrastructure in production.
Model work, APIs, evaluations, and product lessons from Addis AI. Written when there is something concrete to show.
Essays
11
Pieces on language infrastructure, voice, and execution
Themes
- ASR
- Amharic
- Datasets
- Evaluation
- AI Coding
Why This Page Exists
Launch notes and technical essays tied to shipped systems, public tests, and real usage.
More Essays
Good models have to be easy to build with
The Addis AI SDK puts speech, voice, language models, and translation behind one package for Node.js and Python.
Bringing voice into the research workflow
Addis AI and ScholarXIV are bringing voice dictation and spoken research insights into the research workflow.
We gave four voice systems the same Amharic script
The same Amharic customer support script sounds very different across ElevenLabs, Google, OpenAI, and Addis AI.
Addis AI Voices 2 is live
Addis AI Voices 2 ships 28 production-ready voices for Amharic and Afaan Oromo through the platform and API.
15 million API calls changed how we see the market
More than 800 developers made over 15 million Addis AI API calls in ten months. The demand was already there. Language was the gap.
Qwen 3.5 supports Amharic. But is it actually good at it?
We tested Qwen 3.5 on 10,000 real Amharic lines and found the real constraint is not language coverage. It is token cost.
Datasets are the real bottleneck in African AI
Most people talk about models first. In African AI, the real constraint is still the dataset layer.
Why voice AI will dominate emerging markets
In markets where interfaces still exclude large parts of the population, voice is not a feature. It is the product layer.
The gap between demos and deployment
A good demo proves that something can work once. Deployment proves that it keeps working when conditions stop being ideal.
Editorial Note
Dates match the original release or public post. Claims stay close to what we shipped and measured.