Academic & Technical Projects
Academic & Technical Projects
Fine-tuned a DeBERTa-v3-base model with a late-fusion ensemble of linguistic features (syntactic complexity, lexical diversity) to evaluate English essays.
Implemented a co-teaching training paradigm to handle significant label noise and low inter-annotatoragreement (Fleiss’ κ = 0.61), reducing noise impact by 30%.
Achieved a Quadratic Weighted Kappa (QWK) of 0.72, outperforming baseline BERT models and demonstrating robustness in subjective evaluation tasks.
Designed the end-to-end pipeline to handle diverse essay prompts, ensuring scalability and practical deployment readiness.
Technologies: Python, TensorFlow, PyTorch, spaCy, NLTK, Scikit-learn
Designed and implemented a discrete-time PID controller on an Arduino Uno (ATmega328P) for precise motor control and smooth navigation.
Developed an adaptive sensor thresholding algorithm using moving averages to maintain reliability under varying lighting and surface conditions.
Sustained a 125 Hz control loop (8ms period) and achieved stable navigation at 0.4 m/s with seamless transitions between different surfaces.
Validated system performance through rigorous real-world testing, demonstrating consistent operation under environmental uncertainties.
Technologies: C/C++, Arduino, Real-Time Systems, PID Control, Sensor Fusion
Developed a real-time assistive system using Edge-AI to help visually impaired individuals navigate physical environments, emphasizing user privacy through fully local processing.
Engineered a decoupled host-guest architecture: Windows host for camera/TTS handling and QEMU-virtualized
ARM64 Ubuntu guest for isolated AI inference, connected via stable networking.
Implemented low-latency video streaming and AI analysis pipeline using TensorFlow Lite optimized for ARMbased edge devices.
Achieved sub-100ms end-to-end latency for real-time object recognition, enabling practical usability for endusers.
Technologies: Python, TensorFlow Lite, QEMU, WSL2, OpenCV, TTS, Linux/Windows SysAdmin