AI Systems for Humans
I work across natural language processing (NLP), multi-modal AI, and end-to-end ML pipelines.
I care about building AI systems that actually help people — solving real-world problems through reliable modeling, scalable deployment, and measurable performance.
Low-Latency Voice-to-NLP Recommendation System
Fine-Tuned for Fast, High-Accuracy Personalized Search
- Built real-time pipelines integrating Whisper ASR, DistilBERT, RoBERTa spaCy NER, OpenSearch semantic ranking
- Achieved 100% intent accuracy, 99.97% F1 entity extraction
- Delivered recommendations in <0.2s latency
Tensor Decomposition on High-Dimensional Neural Firing Data
Dimensionality Reduction and Feature Extraction
- Applied Tensor Component Analysis (TCA) to PVN spike-train datasets (M.S. Thesis)
- Extracted engagement-linked latent dynamics
- Advanced neural decoding of volitional learning
Embedded AI Deployment in Robotics
Voice-to-Text-to-Robot Motion Command Mapping
- Built embedded AI system for real-time speech-driven robotic motion
- Integrated Julius/SRILM speech recognition with NLP parsing
- Deployed live at 10K+ visitor public exhibition
About
I’m a Machine Learning Engineer with hands-on industry experience in applied AI and a research background in computational neuroscience. My work connects scientific insight with real-world deployment—specializing in natural language processing (NLP), multi-modal AI, and end-to-end ML systems.
In industry, as an AI Robotics Engineer, I developed and deployed an embedded voice-to-motion robotics system showcased to 10K+ exhibition visitors. I implemented the full AI pipeline: speech recognition, real-time command mapping using a fine-tuned NLP model, and robotic motion execution—all running on embedded hardware, with integrated speech synthesis.
In academia, at NYU Grossman School of Medicine, I contributed to neuroscience research under Prof. Robert C. Froemke, analyzing hypothalamic neural activity using Tensor Component Analysis (TCA). My work focused on dimensionality reduction, temporal precision, and oxytocin-linked plasticity—revealing population-level neural dynamics underlying maternal learning.
As a technical initiative, I built a low-latency voice-to-NLP recommender system that transcribes voice input (Whisper ASR), extracts structured intents and entities, and retrieves ranked workout classes—delivering real-time personalized recommendations through a scalable FastAPI and React interface. Intent classification accuracy: 100%. Entity extraction F1-score: 99.97%. End-to-end latency: under 0.2 srconds.
Highlighted Work
- Engineered a real-time AI system that listens, understands, and recommends workouts instantly with <0.2s total latency.
- Fine-tuned transformer models to achieve 97.8% intent classification accuracy and 98.5% entity extraction F1 on noisy ASR-transcribed input.
- Designed an end-to-end scalable backend with OpenSearch boosting, FastAPI, and React for seamless low-latency experience.
Neuroscience Thesis: TCA in Maternal Learning
Statistical modeling for neural decoding & dimensionality reduction
- Modeled PVN spike-train recordings by constructing neuron-time-trial tensors aligned to behavioral events.
- Applied Tensor Component Analysis (TCA) to decompose high-dimensional neural activity into interpretable latent factors across neurons, time, and trials.
- Modeled directional connectivity dynamics with Transfer Entropy, quantifying engagement-driven shifts in neural information flow.
- Demonstrated that volitional engagement sharpens temporal precision, strengthens synchrony, and stabilizes across-trial representations.
Skills
Machine Learning
Frameworks
Algorithms
Neuroscience
MLOps
Languages
Education
Master’s in Computer Science, New York University – Tandon School of Engineering
Thesis: Tensor Component Analysis (TCA) for Neural Plasticity in Maternal Learning
Advisor: Prof. Robert C. Froemke, NYU Langone Health – Neuroscience Institute