The Randstad, Netherlands
With over 6 years of experience, I contribute to transforming customer support at Diabolocom as a Machine Learning Engineer. My work focuses on devising AI solutions in natural language processing, leveraging expertise in natural language generation, statistics, and speech recognition optimization. I design synthetic data generation techniques, build frameworks for uncovering customer insights, and optimize ASR systems for complex conversational data. My mission is to enable organizations to extract actionable insights from text and speech data, enhancing operational efficiency and decision-making. Passionate about fairness and privacy in language models, I bring both academic rigor and industrial impact to solving real-world challenges in AI and NLP.
Working on everything around Agents! From figuring out new way to interact at e-commerce to virtual AI assistant
- Developed trend identification framework, now piloted by 4+ enterprise clients, using CoT prompting and clustering; validated performance with an LLM-as-a-Judge framework. - Currently designing a holistic AI agent evaluation framework (intent recognition, tool-use reliability, task efficiency, RAG), aimed at supporting rapid iteration and safe deployment of enterprise agents. - Boosted NER model's F1 score from 0.65 to 0.85 by generating synthetic training data with Llama 3 70B. - Increased text classification accuracy by 15-30% across 6 benchmarks by developing a ReAct-based agent. - Demonstrated a 2x-4x performance degradation of SOTA ASR models on real-world conversational data (ICASSP 2025). - Led technical pre-sales engagements as a domain expert, designing custom AI solutions to help secure client pilot agreements.
- Proposed a scalable gradient-based algorithm enhancing classifier fairness up to 20% compared to existing methods. Made available as a PyPI Python package, enabling straightforward integration in the existing ML pipeline; published at TMLR, 2023. - Designed and implemented an adversarial learning mechanism leveraging differential privacy to privatize LLM output, successfully obfuscating sensitive attributes in text, and enhancing privacy by over 15% points; published in Findings of EMNLP 2022. - Identified and addressed limitations in existing intersectional fairness evaluation metrics. Introduced and validated, a new generalized metric enabling robust ML model assessments; accepted at EMNLP 2023. - Proposed a novel data generation mechanism improving performance over smaller classes by 60% in unbalanced data settings.
- Orchestrated the development of dialogue systems to query documents ensuring coordination amongst various stakeholders resulting in a general internal platform used across multiple internal projects. - Implemented POC of a multilingual question-answering system, managing end-to-end responsibilities from exploratory data analysis to deployment and monitoring, using tools like Pandas, Docker, and Gradio. - Contributed to various NLP components, including preprocessing module, entity linking, and sentiment analysis.
- Surveyed and benchmarked different reading comprehension systems for semantic search over legal documents. - Co-authored a detailed review on neural network-based Knowledge Graph Question Answering (KGQA) methods and led tutorials on the subject at various forums, sharing expertise and facilitating learning.
- Bootstrapped a large-scale Text to SPARQL dataset using Amazon Mechanical Turk, culminating in a 10x larger dataset. - Refined Transformer architecture to exploit the SPARQL structure, boosting KGQA system F1 score by 8% (published at ISWC).