Nick Mumero

Machine learning engineer in Nairobi, Kenya. I work on speech and language AI for African languages and on LLM systems: training models, building evaluation that holds up, and shipping agents people use.

LinkedIn GitHub Hugging Face mumeronick@gmail.com

Open to new roles from 12 November 2026: machine learning engineer, research engineer or applied scientist; remote, Nairobi, or relocation to Europe, the Middle East or Canada.

Recent work

• Microsoft Research Africa (Aug 2025 – Nov 2026). Speech AI for African languages in Project Gecko. I trained multilingual speech-recognition models from scratch, built text-to-speech with reinforcement-learning post-training, and built the data pipeline and co-developed the evaluation harness behind PazaBench, a public benchmark of 51–52 ASR models across 39 African languages (announcement). I co-authored two papers from this work, listed below; the PazaBench paper was accepted for an oral presentation at Deep Learning Indaba 2026.
• Nexus (Think.ke, 2026 – present). I build the agent layer of Think.ke's entity-intelligence platform, which maps Kenyan public records into a graph of 124K+ entities and 547K+ relationships and checks it for procurement collusion.
• GovBot. Main engineer on the AI assistant for Kenya's eCitizen government-services portal, built at THiNK under the GovStack initiative with GIZ, ITU and ICTA (code).
• Brave Venture Labs (2018 – present). Data scientist, then machine learning engineer. I'm the main engineer on ColdSpring, a job-evaluation and pay-benchmarking system that Brave built for one client, Coldspring, a natural-stone company. It automates the job scoring a compensation consultant did by hand for 30 years: PydanticAI and Claude agents alongside rule-based scoring, in production on AWS, with scores that land within about 2% of the consultant's past evaluations. I also built the AI server behind UnlockAI, a teaching assistant used by Educate's teachers and now reaching more classrooms through changeforce.
• Ecobank Kenya (2023). Led development of Rafiki, the bank's customer-service chatbot, which cut calls to the Kenyan contact centre by 70%.
• Hugging Face. Swahili, Kikuyu and Dholuo speech, translation and dataset releases (nickdee96, thinkKenya).

Open source

• ASR-TTS Paper Daily: a daily list of new speech and language research from arXiv, filtered by rules and tested against labelled examples before each deploy.
• kikuyu-cleaner: cleaning and normalising Kikuyu (Gĩkũyũ) text.
• common-voice-embedding: text embeddings fine-tuned on Common Voice transcriptions for more robust RAG retrieval.

Papers and talks

• PazaBench: Large-Scale ASR Benchmark for Low-Resource Languages in Africa. Mercy Muchai, Nick Mumero Mwangi, Kevin Chege, Samuel Chege Maina. Deep Learning Indaba 2026, oral.
• How Architecture, Typology, and Adaptation Strategies Shape ASR for Low-Resource African Languages: A Controlled Empirical Study. Mercy Muchai, Kevin Chege, Nick Mumero Mwangi, Samuel Chege Maina, Felermino Ali, Elizabeth Ankrah, Ignatius Ezeani. Microsoft Research, 2026.
• Sentimental Agents: Combining Sentiment Analysis and Non-Bayesian Updating for Cooperative Decision-Making (2024, extended abstract).
• IndabaX 2025, Maseno University: building linguistically robust RAG systems with Common Voice transcriptions.
• Data Science Africa 2024, Nyeri: building community chatbots with Rasa.

Tools

Python, PyTorch, Hugging Face, PydanticAI, LlamaIndex, FastAPI, PostgreSQL, Docker, Azure ML, AWS EC2, GitLab CI, GitHub Actions, Prometheus, Grafana.

Education

Bachelor of Technology in Applied Statistics, The Technical University of Kenya.