Hamza Benyamina

— Experience

Where I've worked

1. Nov 2024 — Present
   

   Senior Machine Learning Engineer · Qanoniah

   Riyadh / Makkah

   • Lead the AI direction at Qanoniah: ship every AI feature (legal Q&A, document drafting), pitch performance improvements, and shape feature definition by advising on feasibility, approach, and quality.
   • Built QanoniahAI, a production agentic legal AI platform with modular multi-agent orchestration (6, usable standalone), declarative skills (6, bilingual AR/EN SKILL.md), and tool use; supports deep search (citation-verified) and canvas modes with background streaming and human-in-the-loop; achieves 90% on the legal benchmark.
   • Shipped the document drafter agent: generates Saudi-law contracts using tools and special skills, with block-level and contract-level verifications (compliance, risk, party, and obligation analysis), placeholder handling, and chat-driven edits to specific blocks (no manual selection needed).
   • Built a modular LLMOps layer with multiple providers (OpenAI, Gemini, Groq, on-prem), OpenInference tracing, cost tracking, YAML configuration, and Jinja-templated prompts.
   • Built a full-stack Next.js eval app that scores the agent system across a curated scenario set and gates every release on the result; raised answer accuracy by 15%.
   • Modeled Saudi law as a 1.7M+ node Neo4j knowledge graph; designed indexes and optimized Cypher across three distinct query patterns (ANN, graph traversals, and filtered/text queries); powers hybrid retrieval (vector + keyword + graph) with RRF and reranking.
   • Addressed business blockers to legal-AI adoption: kept all inference in-Kingdom (data residency) and shipped multilingual PII protection (EN, AR) with transfer learning from English into Arabic.
   • Fine-tuned open-source Qwen models for Arabic legal document embedding and reranking; deployed embedding and other models on Triton, vLLM, and HuggingFace TEI.
   • Owned Arabic OCR end to end: continuously benchmarked open-source OCR models, built an evaluation app for legal experts to score outputs, and deployed the winning model in-Kingdom on vLLM/LitServe.
2. Aug 2024 — Feb 2025
   

   Team Lead & Senior AI Engineer (Part-time / Advisory) · Kurasty

   Riyadh

   • Led development of an AI-driven SaaS that drafts government RFPs and auto-audits them for compliance violations.
   • Shipped the first MVP end-to-end: front-end (Next.js), back-end (FastAPI), and AI pipeline — owning all three layers.
   • Set up DevOps and MLOps tooling, configured cloud accounts, and provisioned infrastructure for a fast-moving startup.
   • Hired and onboarded the entire founding team (engineering and operations); standardized workflow on ClickUp, Slack, and Gmail.
3. May 2023 — Nov 2024
   

   Machine Learning Engineer · Alwasaet

   Riyadh

   • Integrated 13+ AI capabilities (OCR, transcription, image recognition, etc.) into a Digital Asset Management platform, automating information extraction from text, audio, and video assets.
   • Built a voice-biometric system generating persistent voice fingerprints and diarizing multi-speaker media for time-stamped attribution.
   • Developed a multimodal RAG agent answering from text, images, audio, video, and YouTube with 90% accuracy globally and per-asset.
   • Delivered an enterprise AI search engine with optimized RAG, reaching 92% faithfulness and 97% relevance on company-asset queries.
   • Orchestrated an event-driven inference pipeline (GCP Pub/Sub, Celery, Redis) and MLOps stack (DVC, Langfuse, OpenTelemetry); accelerated inference 2–4× with ONNX + CTranslate2.
   • Created an AI tutor for 3rd-grade textbooks: advanced Arabic math reasoning, interactive quizzes, in-text help, and a knowledge-graph backbone for complex dependency tracking.
   • Built an LLM-powered course generator producing videos and slides with customizable templates, animations, and natural voiceovers; 4× faster generation.
4. Feb 2022 — 2023
   

   Machine Learning Research Engineer · New York University Abu Dhabi

   Abu Dhabi

   • Developed and published a novel embedding technique enhancing the Tiramisu compiler, reducing MAPE by 3.6% (1 publication).
   • Introduced a novel pre-training method that learns polyhedral geometry, then transfers this knowledge to optimize code performance for complex irregular loops.
   • Engineered an alternative reinforcement-learning method for code optimization on the Tiramisu compiler.
5. Jun 2020 — Dec 2021
   

   Machine Learning Research Intern · New York University Abu Dhabi

   Remote

   • Pre-trained ProBERT, a RoBERTa-based language model, on 6 million source-code files from 14 programming languages.
   • Demonstrated cross-language generalization: ProBERT performed well on programming languages not seen during pre-training.
   • Fine-tuned ProBERT on four downstream tasks (code clone detection, code review, code search, defect detection); surpassed baseline by +2% MRR on code search.
6. Jan 2021 — Mar 2022

   Machine Learning Engineer · Upwork (Freelance)

   Remote · Concurrent

   • Engineered a recommendation engine surfacing peer-generated ideas in a note-taking social network.
   • Applied reinforcement learning for personalized real-time suggestions from user interactions.
   • Built a triplet-based knowledge graph via relation extraction to enrich semantic relevance.
   • Developed an AI-powered EEG/wrist-brain-signal interpreter (research collaboration); achieved 95%+ accuracy and quantized + pruned the model for deployment on a Zynq FPGA.
   • Fine-tuned GPT-2 for synthetic data generation; deployed on a self-hosted GPU server.
   • Completed 21 projects, maintained a 100% job-success rate, and earned the Top-Rated badge.
7. Oct 2019 — Jun 2020

   System Administrator · AIEGS

   Algiers, Algeria

   • Deployed and monitored mission-critical Java services on Apache TomEE; performance-tuned PostgreSQL during peak loads.
   • Integrated the International Postal System with Algérie Post for real-time parcel tracking via XML ingestion into PostgreSQL.