Khalil Ghimaji

Vayetek

Evolving a client CRM in production and modernizing a production application.

• Cut load times by about 5x on the slowest pages of a production CRM (Laravel/Angular): eliminated N+1 queries, slimmed relations, removed an expensive computed attribute.
• Delivered new features from client requests: order synchronization with an external system and automatic catalog generation (Carbone).
• Rebuilt a production application from legacy Laravel/Angular to Next.js and a NestJS REST API, with CI/CD from the first commit (GitLab CI, Ansible, Docker).

Vayetek

Enterprise code-intelligence system for multi-repository codebases.

• Developed AI agents that analyse dependencies across multi-repository codebases, modeled as a knowledge graph and a vector index.
• Built the AST-based analysis pipeline, exposed to the agents through an MCP server, with incremental re-indexing (only modified code is re-analysed).

Elevvo

End-to-end machine learning projects, from raw data to evaluated, production-minded results.

• Predictive maintenance: alert threshold tuned to limit false alarms (3.6% of alerts, 79% of failures detected), served by a containerized FastAPI service (Docker, Pydantic validation, GitHub Actions CI).
• Traffic sign detection: YOLOv8 and a CNN classifier, mAP@0.5 raised from 0.57 to 0.97 at 60 frames per second, ONNX export.
• Walmart sales forecasting: XGBoost with time-ordered validation, error reduced by 22% against the best baseline.

Dental Clinic (confidential client)

Full-stack platform designed and built solo: Flutter and web apps for patients, web platform for clinic staff.

• Implemented role-based access control with hierarchical permissions on the staff platform.
• Protected health data with OPAQUE authentication (NestJS/Next.js): the server never sees the password, neither at registration nor at login.
• Bound access to approved devices through a local reverse proxy written in Rust, so valid credentials alone are not enough.

Shanon Technologies

Open-source desktop alternative to MATLAB/Simulink.

• Contributed to Papaya VL during one development phase: a PyQt6 desktop application where users visually build systems by connecting reusable blocks on a canvas.
• Contributed to translating block diagrams into Arduino-compatible code.

SMARTUP 3S

Invoice processing for highly variable layouts.

• Automated invoice processing with LLaMA Vision where Tesseract OCR was not enough, reaching over 90% extraction accuracy (Python backend, Next.js frontend).

INSAT - National Institute of Applied Science and Technology, Tunis