Mahmut Safa Alıcı

Worked on production AI systems for document, spreadsheet, presentation, and video generation, focusing on evaluation pipelines, agent orchestration, model optimization, and backend reliability.

• Developed and improved AI generation pipelines and agent-based workflows for documents, spreadsheets, and presentations
• Built evaluation and QA mechanisms to detect structural, visual, and content-related generation failures
• Contributed to model selection and optimization, and to the reliability and lifecycle management of sandboxed execution
• Investigated and improved video generation and editing pipelines and long-running backend workflows
• Worked across TypeScript/Node.js backend services, cloud-based infrastructure, and deployment workflows

Completed 500+ deterministic benchmark tasks for agentic AI systems at a sustained pace of 10+ per week, building automated evaluators across nearly a year of active work.

• Designed deterministic benchmark tasks for agentic AI systems across diverse software environments, including LibreOffice and Google Chrome
• Built automated evaluators with JSON schemas and Python logic, testing instruction-evaluator alignment and reducing false-failure paths
• Identified evaluator loopholes and ambiguous success conditions where a successfully completed task could be incorrectly rejected
• Improved evaluator robustness by making success criteria more deterministic

Completed 250+ task-specific LLM development and evaluation assignments across roughly six months of primary active work, spanning coding prompt design, tool/function implementation, and technical model-output evaluation, with a 4.5/5 average platform rating.

• Designed task-specific coding prompts and instructions for model training and evaluation workflows
• Developed assignment-specific callable tools and functions for LLM tool-use scenarios - for example, weather- and calendar-related functions - as individual tasks required
• Evaluated AI-generated code and outputs against defined requirements, correctness, quality, and software-engineering criteria, identifying logical bugs, UI mismatches, and OOP/SOLID violations

Developed a full-stack ERP web application centralizing sales, inventory, and financial tracking.

• Automated KDV calculations, stock validation, and multi-product sales handling
• Designed modal-based UI flows that accelerated daily data entry operations
• Built data visualization features with graphs for actionable business insights
• Extended the ERP into three factory-floor tablet applications for weaving, printing, and garment operations, sharing a common backend and enabling QR-based material traceability across production stages

Arden AI Assistant

An AI layer extending the Arden production-control ecosystem — the ERP plus its connected factory-floor apps — with retrieval over technical documentation and tool-based access to live operational data.

• Converting roughly 8GB of technical documentation into a searchable vector-based knowledge layer for retrieval-augmented responses
• Developing the retrieval and tool-calling architecture that combines document retrieval with structured calls, with Qwen planned as the model layer
• Developing tools for the assistant to query live operational data across the ERP and its connected factory-floor apps — covering weaving, printing, and garment operations

Reel Generation Factory

A multi-stage generative video pipeline that orchestrates short-form content from idea through rendered output - script and scene generation, visuals, voice, captions, and music - for TikTok, Instagram Reels, and YouTube Shorts.

• Orchestrates the full generation pipeline: idea and script generation, scene planning, visual and character generation, voice/TTS, captions, and music
• Supports both Eachlabs and local LLMs per capability, selectable through a provider-selection and orchestration layer built to support additional providers
• Added a partial-run test mode that generates only up to a chosen step (e.g. just the story) at $0 cost, so a local model can be validated without running a full video generation
• Built persistent, resumable workflow state with cost and budget tracking across generation stages
• Renders final short-form video output, with multi-platform publishing planned next

1 / 5