Skip to content
Prestilien.

I love building useful AI and products.

Agents, ML systems and the apps around them, built end to end.

Follow the threadAll work
θ₁θ₂goal
scroll — follow the thread

The thread

Sewn together, one page at a time.

A zine is held together by a single stitch. Scroll and watch the needle pass through each chapter.

  1. Four members of the VALIDE team around a table, two of them wearing VALIDE t-shirts, smiling at the camera
    The VALIDE team

    p. 01

    2021

    Co-founded VALIDE

    Set the technical direction of an education product, from the first architecture sketches to AI-assisted learning.

    Co-founder & CTO

  2. Black-and-white photo of a developer wearing a beanie and headphones, seen from behind, working on two laptops under a sloping ceiling
    Heads down, building for other people

    p. 02

    2023

    Shipped for other people

    Full-stack work at Surfyn: web, mobile and backend features, polished on the outside and dependable underneath.

    Full-stack developer

  3. Three phones showing the Afrik Delices app: the meal planner, the home feed with Ndolè Royal, and the Chef Delice assistant
    Afrik Delices: meal planner, home and Chef Delice

    p. 03

    2023

    Started Afrik Delices

    A food-tech platform for authentic African recipes. I own the product direction and the engineering.

    Founder, product engineer

  4. Hand-drawn agent loop: a user query goes to the agent, which decides whether it needs a tool such as search, web search, a database or an API, observes the result, and loops until it gives a final answer
    The agent loop: decide, call a tool, observe, answer

    p. 04

    2025

    Went deep on production AI

    Agents with explicit control flow, retrieval that cites its sources, ML systems tracked from experiment to container.

    LangGraph · RAG · MLOps

  5. Standing in front of the Multitel building, hands in pockets
    Outside Multitel

    p. 05

    2026

    Research at Multitel

    Reinforcement learning and explainability for intelligent systems, with reproducible experiments.

    AI engineer intern

Field notes

Three things I cannot stop thinking about.

Research · Published paper

Self-supervised and Multilingual Learning Applied to the Wolof, Swahili and Fongbe

Prestilien Djionang Pindoh, Paulin Melatagia Yonta

CRI 2023, 6th Conference on Research in Computer Science · Yaoundé, December 2023

How do you teach a model what speech sounds like when there are almost no transcripts? Let it learn from raw audio first, and pool several African languages so each one helps the others.

  • Self-supervised learning
  • Speech representations
  • Multilingual
  • Low-resource languages
Black-and-white photo of the author presenting at CRI'2023, with a slide about Contrastive Predictive Coding behind him
Presenting the paper at CRI'2023, Yaoundé
Year
2022–2023
Institution
Université de Yaoundé I (UY1)
Supervisor
Paulin Melatagia Yonta
Master's thesis
« Apprentissage multilingue et autosupervisé de la représentation de la parole pour les langues peu dotées »
waveformf(θ)no labels needed123

Schematic. Each shape stands for one language.

Fig. 4

The method, in one picture

  1. Listen without labels

    Raw audio only: CPC, wav2vec and a bidirectional CPC learn representations without any transcripts.

  2. Pool the languages

    One encoder is trained on Wolof, Swahili and Fongbe together, using the ALFFA speech corpora.

  3. Check what it learned

    The encoder is frozen and a DeepSpeech-like recogniser is trained on top, one language at a time.

Selected work

A few things, chosen on purpose.

Agents, ML and products, built end to end.

Show me

3 projects shown

Interested? There is more — open the archive

Certifications

The paperwork, framed.

Credentials I have earned, each one verifiable.