Technical AI Research Associate at Keywords Studios (Amazon Nova Act) (2025-07 – 2026-08)
Owned evaluation data, quality metrics, and failure analysis for Amazon's production web automation agent.
- Analyzed 1,000+ agent interaction logs and categorized them into a failure taxonomy of 150+ documented cases, giving engineering a prioritized fix list that improved issue resolution by 35%.
- Defined measurable success criteria for 60+ task types and implemented them as automated verifier code, replacing subjective review with reproducible pass/fail labels.
- Found a measurement flaw where outcome-only checks recorded false passes when the agent skipped a required step, and fixed the metric to validate intermediate actions as well as end states.
- Monitored label quality across 500+ weekly interactions, leading a 5 person annotation team through spot checks and calibration syncs to keep reviewer judgments consistent.
- Built ground truth datasets from 250+ structured prompts and rubric based scoring criteria, improving model response accuracy by 18%.
Machine Learning Scientist at EntreVita Inc. (2024-07 – 2024-12)
Built and deployed predictive models for an AI precision nutrition product, from prototype to production.
- Built a YOLOv5 and MiDaS pipeline estimating real world height from a single image, validated at 96% detection accuracy on labeled test sets.
- Fine tuned LLMs with QLoRA for a dietician recommendation pipeline, improving nutrient match relevance by 22%.
- Containerized and deployed inference services with Docker on Azure, cutting serving latency by 40% while holding 99% uptime.
Senior Executive (Data Engineering) at Vodafone Intelligent Solutions (2021-08 – 2023-03)
Promoted from Graduate Engineer Trainee in Aug 2022. Built data pipelines and models feeding business reporting.
- Built a GCP pipeline moving raw data through Cloud Storage and Vertex AI into BigQuery, delivering analysis ready tables to downstream reporting teams.
- Built ML based anomaly detection that flagged data quality issues before they reached stakeholder facing reports, and automated recurring ETL jobs, improving data quality metrics by 37%.
- Shipped a TensorFlow personalization model that increased customer engagement by 27%.
