AI/ML engineer with 6 years shipping intelligent systems end-to-end, and founder of Sefar AI — where I take agentic, retrieval, and vision systems from prototype to production. MSc in Artificial Intelligence and BSc in Computer Science (USTO, Oran). I build AI the way you build infrastructure: bounded, typed, evaluated, and observable.
If it can't be tested, traced, and constrained — it's not ready.
| Domain | What I build |
|---|---|
| ⛓️ Agentic Systems | Tool-calling agents with schema validation + strict allowlists · plan→execute pipelines with checkpoints · multi-agent workflow automation · explicit memory/state · guardrails & rollback paths |
| 🔎 Retrieval (RAG) | Hybrid retrieval (BM25 + dense) · adaptive/semantic chunking · reranking & evidence packaging · citation-enforced generation · eval-gated regression (Ragas / DeepEval / custom) |
| 👁️ Vision & Multimodal | Detection / segmentation / OCR pipelines · GPU batching & latency tuning · raw outputs → reliable structured data · document-heavy perception stacks |
| 🗣️ Voice & Speech | Voice cloning & TTS (XTTS) · real-time voice agents · speech pipelines wired into agentic backends |
| 📊 Structured ML | Gradient boosting (XGBoost / LightGBM / CatBoost) · explainability-first (SHAP / LIME) · drift-aware retraining |
| Project | What it is |
|---|---|
| glove-from-scratch | GloVe word embeddings implemented from raw PyTorch — co-occurrence matrix, weighted least-squares loss, no gensim, no pre-trained vectors. |
| word2vec-from-scratch | Word2Vec from scratch — Skip-Gram, Negative Sampling & CBOW, fully commented, no shortcuts. |
| Multi-Agent-Healthcare-Assistant | Coordinated multi-agent system applying tool-calling + retrieval to a high-stakes domain. |
| XTTS_Voice_Cloner | Voice cloning / multimodal speech synthesis pipeline built on XTTS. |
| vehicle-vision-system | End-to-end computer-vision pipeline turning raw detections into structured, actionable data. |
| 🔒 Sefar AI — production work | Agentic backends (FastAPI), Next.js frontends, self-hosted deployment & CI/CD. Private. |
Languages
Deep Learning & ML
LLM / Agents / RAG
Serving, Infra & Cloud
- Constrain before you scale — tools, schemas, and policies first.
- Make state explicit — memory, plans, and intermediate artifacts are first-class.
- Measure grounding over persuasion — no eval, no deploy.
- Instrument everything & optimize for p95 — traces, metrics, latency budgets, cost as a hard constraint.