Chima Emmanuel
Hi, I'm Chima. I train models and dabble between computer vision, speech and computational biology. On the speech side, I mostly work on the languages people in Nigeria actually speak, which tend to get left out of the big labs' datasets.
I run Modular, a small research company. Everything we train goes out openly onHugging Face: a text-to-speech model and a speech recognizer that handle switching between English, Pidgin, Igbo, Yoruba and Hausa mid-sentence, a 2.4M-parameter time-series forecaster trained from scratch, a protein stability predictor, and a few others. Almost all of it was trained on GPUs rented by the hour, so I've learned to be stingy with experiments and honest about results.
I do SWE at Federal University of Technology Owerri, and I contribute to open-source ML infrastructure, most recently PyTorch's ExecuTorch.
This site is where I write things down: what I'm building, what worked, and especially what didn't.
Recent
- Released Dynamic ASR, speech recognition for Nigerian code-switched speech, fine-tuned from Meta's Omnilingual ASR.
- Released Dynamic Accent, one voice that speaks five languages and switches between them inside a sentence.
- Trained Modular-TS-2.4M, a tiny forecasting foundation model that runs on a CPU.
- Wrote Building R&D on Soft Ground, an essay on doing research where the infrastructure is missing, and The Local Minimum, a short book on what failed experiments taught me.
- Co-authored a journal paper on an AI app for poultry disease diagnosis.
Writing
Autoencoders and Representation Learning in Vision
Autoencoders are a type of neural network that compress data into a lower-dimensional space and then reconstruct the original input from that compressed representation.
The Weird Geometry of High-Dimensional Representations: Why Your ML Models Behave Differently
In high dimensions, our geometric intuition fails. All points become equidistant, random vectors become orthogonal, and the very concept of a nearest neighbor collapses.
Why Neural Networks Fail on Constraints and why Lagrangian Geometry works better
Exploring why standard neural networks struggle with strict constraints and how Lagrangian geometry provides a more principled approach to obedience rather than approximation.
Licensing & Contact
Code is released under the MIT License. Written articles are licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC-BY-NC-SA).