Sara Hwanjiru

Contributed to image generation evaluations by reviewing how well generated images follow prompts, focusing on quality and safety. Performed audio transcription and ASR labeling to support large-scale training of speech recognition models. Carried out speaker diarization tasks by identifying and labeling multiple speakers in audio conversations for voice assistant use cases. Provided verbatim and ASR-corrected transcriptions for phone calls with over 98% accuracy, ensuring transcripts are reliable for downstream tasks.

Rated AI model responses on helpfulness, harmlessness, and honesty (RLHF) across large-scale LLM training datasets. Caught and documented hallucinations in AI-generated text, pairing each failure with reference-backed corrections. Evaluated safety and policy adherence of model outputs, detecting unsafe, biased, or non-compliant responses according to platform policies. Assessed code generation responses across multiple programming languages for correctness, efficiency, and style. Conducted response ranking and Likert-scale evaluations to inform reinforcement learning and shape model preferences.

Labeled video datasets for action recognition, segmentation, and labeling of human activities and interactions to support training of action detection models. Segmented long videos with scene and event labels for video surveillance and sports video analytics. Classified short-form text data (e.g., social media posts, reviews, and headlines) for sentiment and category. Annotated Named Entity Recognition (NER) and Part-of-Speech (POS) tasks on multilingual datasets to support training for search engines and conversational AI. Created image classification labels and generated tight bounding boxes to train computer vision models for retail and e-commerce applications.