How work transforms. How systems stay accountable. What becomes possible.
Three directions — how people adapt, how systems behave, what becomes possible.
Not training on tools. Redesigning work around new capabilities.
Accountable systems at scale. Safety isn't theoretical — it's how we build.
Closing the gap between demos and deployment.
Research across our three focus areas.
Production dialogue systems face a critical challenge: achieving high accuracy while maintaining low latency at scale. This work introduces Symbol Tuning and C-LARA — two complementary approaches that enable enterprise deployment of LLM-powered intent classification at a fraction of the computational cost.
Chain-of-Intent combines Hidden Markov Models with LLMs to generate context-aware dialogues through self-play, addressing the fundamental data scarcity problem in conversational AI.
Combines a fine-tuned compact model with retrieval-augmented LLM architecture for cross-lingual intent classification. Achieves 3.67% accuracy improvement across six languages.
Using real university admissions data, we challenge the assumption that fairness and accuracy exist in tension. Our findings reveal that ML models exceed human fairness consistency by 14-18% while maintaining comparable accuracy.
Combining Byte-Pair Encoding with gated multi-head hierarchical attention for nuanced assessment of semi-structured data. Achieves F1-score of 0.8453 while offering interpretability for high-stakes decisions.