AI Engineer · Full-Stack Developer · Dakar, Senegal
Co-founder of zeroRL · Building and running AI systems at DIT
- ~3 years building ML systems and full-stack apps, including production deployments
- MSc Artificial Intelligence, Dakar Institute of Technology (Very Good, class valedictorian)
- Teach Python, data collection, and deep learning
- Open to roles in AI / ML engineering and full-stack development
Production facial-recognition attendance system used on campus and wired into MyDIT.
Pipeline
- Detection: RetinaFace · embeddings: ArcFace (r100) via InsightFace / ONNX Runtime (CUDA; dlib fallback)
- Continuous multi-camera ingest (USB DirectShow + RTSP through FFmpeg)
- Gallery match with cosine/threshold scoring and multi-frame confirmation before logging an event
- Concurrent multi-person tracking; unknown faces kept as visitors; local enrollment when needed
- One structured attendance record per person/day (arrival fixed, departure updated live)
Systems
- Flask + Waitress admin UI · PostgreSQL in prod (SQLite for local) · optional InfluxDB latency/throughput metrics
- REST/JWT sync of student photos and embeddings with MyDIT
Research (papers in progress)
- Spatio-temporal patterns from the same streams (schedules, co-presence, cross-camera transitions)
- Clustering recurrent unknowns into stable identities
- Calibrated anomaly alerts + continuous-learning loop with a reproducible eval protocol
Repo: Silanoa/model_pointage
zeroRL is a full PyTorch reinforcement-learning framework built for experiment control — not a thin wrapper around a single algo.
You can work at three levels of abstraction:
- High-level —
easy_train_ppofor a complete PPO run (agent, vectorized envs, buffer, logging) in a few lines - Mid-level —
BaseTrainwith swappable agent, env, buffer, and update function - Low-level — primitives to assemble the training loop yourself (
Buffer, GAE/PPO ops, logging, env helpers)
Includes Gymnasium integration, custom MuJoCo environments, config system (TrainConfig / AlgoConfig), TensorBoard / W&B tracking, checkpointing, and a growing example suite. On PyPI: zerorl (Apache-2.0).
I contribute across the stack: training pipeline, envs & examples, Windows support, logging/export, and CI.
University platform (admin + students) I helped take to production: Laravel modules, integrations, deploy — including connecting SmartAttendance to campus workflows.
Built for courses I teach (IoT / data / ML practice):
| Repo | Focus |
|---|---|
| projet1-dht22-thingsboard- | ESP32 DHT22 → ThingsBoard (MQTT) + PC simulation |
| projet2-pompe-meteo-ml | Greenhouse: ESP32 pump + weather ML agent + ThingsBoard |
| projet3-iot-telegram | IoT alerts via Telegram bot |
Hotel ops chatbot (booking, services, cancellations): PyTorch intent model, NLTK, MySQL, plus an RL loop to improve replies over time — Silanoa/hotel-chatbot.
ML / DL · Computer Vision · NLP · RL · data pipelines
- Extending SmartAttendance toward adaptive video analytics and writing papers on that track
- Pushing zeroRL further — APIs, algorithms, envs, developer experience
- Actively exploring opportunities in AI / ML and full-stack



