DialFlow — Gen‑AI Voice Agent (Winner, 2025)

• Award: 1st place, First Moroccan Gen‑AI Voice Bot Hackathon by AI Crafters @ Technopark (MITC), Casablanca
• Mission: customizable AI voice agent for multilingual customer service with task automation and analytics
• Stack: Twilio (voice/telephony), FastAPI (real‑time API), Redis (state/caching/queues), ElevenLabs (TTS), LangChain (prompt/flow orchestration)
• Architecture:
  • Voice call ingestion via Twilio; secure webhooks and call session management
  • Streaming pipeline: FastAPI WebSocket endpoints with back‑pressure; worker isolation for TTS/STT tasks
  • State and rate control via Redis (session state, throttling, caching of prompt/tool results)
  • Orchestrated dialogs using LangChain tools/agents; deterministic prompt versioning and audit
  • Observability: latency percentiles, error budgets, call completion funnels, per‑intent success metrics
• Reliability & scale: adaptive batching for synthesis, retry policies and circuit breakers, canary releases for new prompt versions
• Value: production‑ready voice assistant foundation with measurable service KPIs and cost controls; accelerates integration for Moroccan businesses

Forensic Image Feature Matching — DICE | DATALAB

• Scope: designed and operated a production model‑serving app for forensic image feature matching with online quality monitoring
• Stack: FastAPI/BentoML, Docker/Docker Compose, MLflow, KS‑based drift detection, separate ML workers and web workers for concurrency; Kubernetes/ECS‑ready
• Key challenges solved:
  • Scalable concurrency: isolated ML workers for heavy inference, web workers for HTTP handling; tunable worker pools and adaptive batching
  • Reproducibility: MLflow‑tracked runs with pinned data/model/code; deterministic preprocessing
  • Online monitoring: KS scores on match distributions and response latency; automated alerts and rollback strategies
  • Retrainable workflows: templated pipelines for data updates and model refresh with CI/CD gates
• Value: improved robustness under real‑world distribution shift, predictable latency, and safer deployments via progressive delivery

Child Educational Platform using SADAI/Allam LLM in IBM Cloud — Atlas Innovators (Finals, Riyadh 2024)

• Event: Allam Challenge 2024 Finals, Riyadh (Sep 7–10); team Atlas Innovators — Othman Moussaoui, Ismail Hamdach, Hicham Maghraoui, Ibrahim Lahlou
• Mission: interactive Arabic learning platform for children; enable Allam LLM in IBM WatsonX to “see”, “hear”, “speak”, and “generate images”
• Backend role:
  • FastAPI service exposing clean endpoints for model inference and multimodal interactions
  • Child/parent management: entity modeling, authentication/authorization, and session handling
  • MySQL storage for profiles and insights (engagement, progress, comprehension metrics) to personalize learning
  • Integration as the final stage: endpoint governance, schema contracts, and reliability checks before demo
• Architecture: modular services for perception (vision/audio), TTS/ASR, and LLM reasoning; deterministic prompt versions and telemetry for education KPIs
  • Components:
    • Vector store and document embeddings for semantic search over educational content
    • LLM serving backend with Allam and open‑source models: Whisper Large v3 (ASR), XTTS (TTS), FLUX.1 (text‑to‑image), vision modules
    • Providers for inference acceleration and model variety (e.g., Groq)
    • Database: MySQL for child/parent profiles and insight metrics
    • Profiling backend: Node.js and FastAPI integration for endpoint management and telemetry
    • Frontend: Axios and fetch bridge to modules (Story, Chatbot, Parent Interface, Child Interface, Authentication, Quiz)
• Value: production‑aligned backend foundation that supports safe, measurable, and adaptive learning experiences for kids

Smart NLP Pipeline (AWS ECS) — 3D Smart Factory

• Built an end‑to‑end text processing pipeline using AWS S3 for storage and SymSpell for normalization and retraining
• Served predictions via Flask + WSGI; containerized and deployed on AWS ECS with production configuration
• Introduced data quality checks and automated QoS thresholds across multiple iterations
• Value: reliable retraining loop with traceable artifacts and stable inference in production