Joy Dasgupta
I am an AI Evaluation and Alignment Specialist with more than five years of experience working across RLHF, safety red-teaming, prompt engineering, multimodal computer vision, and training-data quality assurance.
About
I am an AI Evaluation and Alignment Specialist with more than five years of experience working across RLHF, safety red-teaming, prompt engineering, multimodal computer vision, and training-data quality assurance. I have evaluated reasoning, factuality, safety, and policy adherence in large language models, while also building and validating robotics and gesture-recognition datasets. My work with Outlier, Welo Data, Appen, Lionbridge, and independent AI projects has helped me maintain consistently high quality scores and deliver practical feedback for model improvement.
I enjoy turning complex evaluation requirements into clear rubrics, reliable workflows, and actionable insights. I am especially interested in agentic reasoning, multimodal perception, hallucination detection, and responsible AI development. Alongside my technical work, I am a recognized vocalist and YouTube content creator, which gives me an additional creative perspective on audio, video, and digital communication.
- 10 years' experience
- Joined October 2026
Skills· 11
Languages
- EnglishFluent
- BengaliFluent
- HindiAdvanced
Industry experience
- Software & Internet
- Computers & Electronics
- Media & Entertainment
- Professional Services
- Travel & Hospitality
- Telecommunications
Work experience· 7 roles
Trained generative video and image models for a global social media platform. Evaluated visual outputs for realism, compositional coherence, artifact distortion, and safety compliance, providing quality feedback to support iterative model improvements.
Performed high-volume relevance rating, query-intent mapping, and commercial-policy compliance audits with sustained accuracy above 98%. Produced qualitative feedback to improve search-query alignment and resolve semantic ambiguities in retrieval pipelines. Flagged navigational-intent mismatches, regulatory non-compliance, and deceptive commercial content.
Engineered and scored more than 2,000 benchmark prompts covering reasoning trajectories, mathematical logic, factual accuracy, and safety constraints across large language models. Designed evaluation rubrics that improved cross-reviewer agreement and consensus scores by 12%. Conducted adversarial red-teaming to identify hallucinations, prompt-injection vulnerabilities, and policy breaches, and delivered detailed RLHF feedback to support model alignment and preference optimization.
Education
Certifications and awards
Hire Joy Dasgupta today
Post a job and invite Joy, or contact them directly. No hidden commission.