Vision-Language Researcher (MSc / Phd) - Evals

Duzz.ai is building autonomous QA verification agents for mobile games and AI-generated software. Our agents plan and play long game sessions, ground UI elements in images and video, and answer questions about what happened on screen. We're looking for an MSc/PhD student to join the Algorithms & Agents team and own the evals infrastructure that keeps our vision-language stack honest.

What You'll Do

• Build and scale the evaluation infrastructure for our agent and perception stack
• Define evals for our agents capabilities: UI grounding, long video VQA, long/short-term planning.
• Curate datasets, design benchmarks, and run systematic comparisons across models and harness strategies.
• Analyze agentic decision trajectories and feed insights back into our stack.
• Work closely with the agents team to ship eval-driven improvements

Requirements

• MSc / PhD student in CS, EE, or related (active enrollment)
• Strong Python and ML fundamentals
• Hands-on experience with at least one of: VLMs, video understanding, grounding, VQA, or LLM evaluation
• Comfortable reading recent papers and turning them into runnable experiments
• Ownership mindset - happy to define the problem, not just solve it

Nice to Have

• Experience with long-form video models (Gemini, frame-sequence reasoning, temporal grounding)
• Prior work on eval frameworks (Langfuse, lm-eval-harness, custom harnesses)
• Mobile UI / GUI grounding background

Why Duzz.ai

• Real research problems with direct product impact. Evals here drive what ships
• Early-stage, small team, fast iteration cycles
• Access to a unique dataset: long-form mobile gameplay video at scale