Sɔʼ pəpúŋ“Welcome” in Ghomálá', a Bamiléké language of West Cameroon
AI engineer and researcher, from models to production systems.
I build AI models, data pipelines and the backends that put them in production, with research in NLP and optimal transport on embedding spaces, and work on the languages of Cameroon.
AI engineer and researcher working across the full life cycle of AI systems, from theory to production. My core research is natural language processing through the lens of optimal transport: Wasserstein distances, spectral perturbation theory and eigenspace analysis to measure how multilingual embedding spaces move across models and source languages (Master's research, UMMISCO lab, IRD / University of Yaoundé 1).
On the applied side, I design LLM systems that reason and verify their answers: our few-shot text-to-SPARQL pipeline won the Spanish with-mentions track of the WikiKGQA challenge at ISWC 2026 and is published in CEUR-WS. I also take models to production for the languages and accents of Cameroon (Ghomálá' translation, Cameroonian-French speech recognition).
And I build the distributed backends that carry them: I maintained and evolved a real-time platform of 30+ event-driven microservices (RabbitMQ, Redis, Traefik) fed by sports odds providers, integrated Mobile Money payments and third-party APIs on NestJS SaaS products, and ship everything with Docker and CI/CD.
Curious by nature, and open to neighbouring research questions.
What I use, grouped by domain, and what I do with it.
Python
TypeScript
JavaScript
SQL
R
PHP
PyTorch
scikit-learn
TensorFlow
Hugging Face Transformers
NumPy
pandas
SciPy
Jupyter
LangChain
LangGraph
LlamaIndex
CrewAI
Model Context ProtocolHugging FaceTool callingPrompt engineeringRAGFAISSChromapgvectorSPARQLPostgreSQL
MariaDB
MySQL
Redis
RabbitMQ
ScrapyPentahoETL pipelines
NestJS
Node.js
Django REST Framework
Flask
Next.js
Symfony
Prisma
OpenAPI
JWT / OAuth 2.0Power BIR Shiny
PlotlyMatplotlib
React dashboards
Gradio
Docker
Traefik
GitHub Actions
Linux
Let's EncryptHugging Face SpacesEach project sits in the domain it illustrates. Open a card for screenshots and technical details.
Models are only useful once people can use them. I take them from the notebook to a running service.
Feature engineering, model selection and evaluation, deep learning with PyTorch and TensorFlow, model serving behind REST APIs. Neural machine translation, speech recognition and recommender systems taken to production.
Neural machine translation for an under-resourced language: French to Ghomálá', a tonal Bamiléké language of West Cameroon, backed by a sourced dictionary and an open contribution platform.
Speech recognition that understands Cameroonian accents: Wav2Vec 2.0 models fine-tuned on recordings of speakers from Yaoundé, live on Hugging Face.
Design and production deployment of a recommendation model integrated into the gHealth Village Connect platform.
My core research is NLP through the lens of optimal transport: measuring how meaning moves between embedding spaces, models and languages.
Optimal transport for NLP: Wasserstein distances between embedding distributions, spectral perturbation theory and eigenspace analysis to quantify how multilingual representations shift across models and source languages. Applied side: LLM-based semantic parsing and knowledge-graph question answering over Wikidata.
My main research line: how much does the meaning of a corpus move when it is embedded by different multilingual models or through different source languages? Optimal transport and spectral perturbation theory give a principled answer.
LLM-based semantic parsing that turns natural-language questions into SPARQL queries executable on Wikidata. Published in CEUR-WS.
Benchmarks keep me honest: regular data science and AI competitions on Codabench, Kaggle and Zindi, and an active member of the KmerAI community.
I design LLM systems that reason, call tools and check their own answers instead of guessing.
Agent orchestration with LangChain, LangGraph and CrewAI, tool calling and Model Context Protocol servers, retrieval-augmented generation with LlamaIndex and vector stores (FAISS, Chroma, pgvector), multi-candidate sampling with execution-guided selection, open-weight model inference.
A generate–execute–select loop: the LLM writes candidate queries, runs them as a tool against the Wikidata endpoint and keeps the one that returns a valid answer.
A series of LLM application projects: chains, tool-using agents, state graphs and RAG pipelines.
Reliable data comes first. I collect it, clean it, move it and turn it into decisions.
ETL pipelines, large-scale web scraping, relational data modelling, ingestion and runtime-data capture, message-based data flows and caching layers. Exploratory and statistical analysis, SQL analytics, interactive dashboards (Power BI, R Shiny, React) and decision-oriented reporting.
Automated scraping system that builds corpora of biblical texts for the lab's natural language processing work.
Scrapers for the main Cameroonian job boards feeding an aggregator that ranks offers against the user's CV.
End-to-end vehicle data traceability: Python capture scripts, Pentaho-orchestrated ETL jobs and a live monitoring dashboard.
Dozens of projects on real data, including Products Sales Strategy, Android App Market, Predicting Credit Card Approvals and Nobel Prize Winners.
I build and maintain the distributed backends that carry real traffic, payments and third-party integrations.
Microservices and event-driven architectures (RabbitMQ), distributed caching (Redis), authentication and authorisation (JWT, OAuth 2.0), payment gateways and third-party API integration, OpenAPI documentation.
WhatsApp becomes a 24/7 booking desk: an AI assistant handles the conversation, the slot, the Mobile Money payment and the reminders.
E-learning platform simulating professional situations for the BTS Tourism diploma: a virtual travel agency with mock exams.
Daily stock and sales management for a bar: deliveries and counts in, sales, margins and profit computed automatically.
Maintaining and extending a transactional platform of more than 30 microservices, fed in real time by odds and market providers (FeedConstruct, API-Sports).
My core research on optimal transport for NLP, and a peer-reviewed paper on knowledge-graph question answering.
Emeric Cyrille Kenmegne Fokam
University of Yaoundé 1
How much does a representation change when the same corpus is embedded with different multilingual models or source languages? This work models the geometric transformations of embedding spaces with spectral perturbation theory, and measures distances between representation distributions with Wasserstein optimal transport and eigenspace analysis.
Emeric Cyrille Kenmegne Fokam, Gilbert Ebenezer Tela Tela
WikiKGQA 2026, co-located with ISWC 2026 · CEUR Workshop Proceedings, Vol. 4284
Winner of the Spanish with-mentions track
We report on our participation in the two with-mentions tracks (English and Spanish) of the WikiKGQA @ ISWC 2026 challenge, whose task is to translate a natural-language question into a SPARQL query executable against a Wikidata subgraph. Our system chains four stages: a TF-IDF retriever that selects three similar training examples, a few-shot prompt sent to an open-weight LLM, generation of five candidates with mixed temperatures (one deterministic, four at 0.7), and selection of the first candidate returning a non-empty result on the challenge endpoint. An error analysis isolates two resistant categories: type-filter questions, which rely on the P31/P279* chain that is incomplete on the endpoint, and COUNT queries, for which the LLM frequently forgets the aggregate. The pipeline uses only open-weight models (gpt-oss-120b, llama-3.3-70b) served free of charge through Groq Cloud.
My role: Paper presentation · WikiKGQA challenge track winner
ISWC is the premier international venue for research on the Semantic Web, knowledge graphs and linked data. The 2026 edition hosts the first WikiKGQA challenge, which evaluates question-answering systems over Wikidata in English and Spanish. I present our text-to-SPARQL system there, winner of the Spanish with-mentions track.
My role: Organising committee, logistics
International workshop of the African Society in Digital Sciences (ASDS), run by the AI4AfricanLanguages initiative with support from the French National Research Agency. It brings together researchers working on digital tools for under-resourced African languages. I was part of the organising team, in charge of logistics.
Two threads that feed each other: building systems in industry, and research at the university.
Freelance · Remote (France)
A first client project: database design in MariaDB, interface in SASS and Bootstrap, deployment on Platform.sh.
HS4HEALTH · Yaoundé
My first model in production: a recommender for gHealth Village Connect, from data cleaning to REST APIs.
Africa Systems · Douala
Data at scale: vehicle data capture and traceability, collection jobs orchestrated with Pentaho, Flask APIs and a monitoring dashboard.
Core Extreme · Remote
Microservices and payments: secure authentication (JWT, OAuth), PostgreSQL, payment gateways, on products such as ReserveIT and a sunbed booking app.
AgdTech · Yaoundé
Event-driven architecture at scale: maintaining and extending a transactional platform of more than 30 microservices fed by real-time odds providers.
Lycée bilingue de Tougang 2 · West Cameroon
Mathematics and physical sciences: the foundations of everything that follows.
University of Yaoundé 1 · Yaoundé
Five years of computer science, statistics and machine learning, completed with the DataCamp Data Scientist Professional track and four certifications.
UMMISCO · Yaoundé
Into research: optimal transport and spectral perturbation theory to compare embedding spaces. In May, on the organising team of the DASSA'25 workshop on African low-resource languages.
A need, a question or a collaboration: write to me, I reply within 24 hours.