• Developed an on‑device (Android, Snapdragon 765G) inference pipeline for video processing and assembly using a visual‑language model. Leveraged PyTorch Quantization and PyTorch Mobile to achieve approximately 3× lower peak memory consumption.
• Refactored the visual‑language model to support statically‑typed forward passes and data‑dependent control flows, reducing inference latency by 16.67%. Additionally optimized memory consumption through operator fusion and parameter hoisting techniques.

• Achieved an increase of 4.74 units for Rouge score and 3.60% for Accuracy@1 improvements for few‑shot learning in Flan‑T5 transformer, by expanding their capacity to be able to process more in‑context example within the same context window length through sub‑batching.
• Inched closer to finetuning‑like performance through pure in‑context learning (ICL) by 2.16 units of Rouge score and 3% for Accuracy@1 through mesa‑optimization where the transformer acts like an optimizer itself during inference.
• Improved the cross‑domain transfer capabilities of a transformer (Flan‑T5) by 1.68 units for Rouge score and 1.3% for Accuracy@1 through incorporating both cross‑ and within domain question‑answer samples within a limited context window length of 512 tokens.
• Evaluated and verified the effectiveness of both sub‑batched ICL and mesa‑optimization during inference on both Adobe and public datasets.

• Built a federated solution of personalized recommendation systems and classifiers for real-time on-device learning, by using early-stopping for client-side updates and drift adaptation at server-side, to achieve robustness against concept drift (distribution change with respect to time).
• Presented a drift-aware adaptive optimization strategy that can quickly adapt to various concept drift patterns (sudden, incremental, and recurrent), by taking into account historical gradient updates and identifying change in gradient magnitude as drifts, to achieve lowest accuracy drop and fastest recovery from the said drifts.
• Evaluated the proposed algorithm on benchmark computer vision and natural language processing tasks, achieving the lowest accuracy dip difference (the lower, the better) of 1.48%-2.99%, while the best performing baselines exhibit 3.15%-9.22%.

• Led technical sessions in a 6 weeks workshop program on applied deep learning, as a head mentor of 50+ students, by daily presenting and teaching deep learning concepts.
• Developed, and presented machine learning and deep learning curriculum to high-school and college students everyday
• Collaborated with computational culture researcher to identify patterns interesting for Learning Sciences to learn how to make AI more accessible to younger, diverse group of students
• Supported, and un-blocked 80+ students and 15+ mentors in learning program material through individual and group discussions, office hours, mentoring stand-up hours
• Contributed in curriculum building to support the daily discussion sessions on the nuances of applied deep learning concepts like optimization, generative networks, algorithmic biases, regularization.
• Managed teams of 5 in multiple SureStart programs and guided the teams to build a deep learning based capstone project addressing real-world challenges like awareness on harmful ingredients in processed food (Winner in Feb 2023), marine pollution (Runner up in Jun 2022), automotive safety (Winner in Feb 2021), and climate change.

• Led a team of 4 and won "Virtual AI Learning Make-a-thon" for creating driveAId, an in-car monitoring tool used to improve teens' driving ability
• Built the driveAId system with 2 models: facial experssion recognition and gesture recognition to detect distracted and perturbed driver cases
• Addressed the ethical concerns related to teen driver monitoring system by making it non-invasive, transparent, and local-to-the-device
• Mentored trainees in Machine Learning skills, while also applying the taught skills to real-life applications having ethical concerns

• Developed a prototype for an “Emotion-enabled” smart fridge, detecting the available food inside a refrigerator, and the mood of the user with the "Affectiva Facial Expression Recognition SDK" to suggest the food intake accordingly
• Trained the VGG-16 food classifier on Freiburg Grocery dataset with the accuracy of 76.16%
• Implemented a multiclass incremental learner classifier decision tree which correlates the mood of the user and what food they are picking up from the fridge
• Became familiar with the market analysis, pitching the project idea, patent creation and honed the technical aspects of data acquisition, data synthesis, affect analysis, and personalization from facial analysis

• Created the simulations and visualizations of inner workings of operating systems and processors in Java
• Utilized JNLP [Java Network Launch Protocol] to make the Java apps render on client-side while using server resources with ~1000 simultaneous connections capacity
• Refactored the database of the learning platform site on which these simulations were embedded, with 60% increase in speed-up efficiency
• Provided a mobile interface for the site which had inherently no support for mobile version of servlet render