Aug 2026 – Present
AI Engineering Intern
DrRx.ai
Retrieval-Augmented Generation
- Developing a RAG pipeline leveraging BM25S retrieval and Qwen2.5-7B-Instruct over 2K+ pharmaceutical documents, integrating data from PostgreSQL and AWS S3.
Industry Experience
My industry work focuses on large-scale machine learning systems, geospatial modeling, and production-ready AI experimentation.
May 2026 – Oct 2026
AI Fellow
AI4ALL Ignite Summer Accelerator
Computer Vision & Assistive Technology
- Developed an end-to-end Braille-to-speech computer vision pipeline integrating YOLO11n, EfficientNet-B0, Liblouis, and Piper TTS, achieving 96.15% detection mAP50, 99.41% classification accuracy, and 97.32% character coverage.
- Trained and evaluated detection and classification models across 4 datasets and 5 model families, scaling to 10,117 augmented images and 193K+ character crops, and performed end-to-end error analysis to identify false-positive detection and reading-order failures.
May 2026 – Aug 2026
Machine Learning Intern
Regeneron
QA Validations & Operations
- Deployed a Streamlit app for the QA Validation and Operations team after evaluating 10+ ML approaches and selecting a TabPFN Regressor.
- Engineered 40 trend-based features from 500K+ GMP manufacturing records to predict remaining useful life (RUL) and classify equipment-breakdown risk using an RUL threshold.
- Implemented K-fold cross-validation with single-class fold rejection and data-leakage safeguards, achieving 67.8% ROC-AUC.
May 2026 – Aug 2026
Google Summer of Code 2026 Contributor
DeepChem · Google Summer of Code
Molecular Language Models
- Selected as 1 of 5 from 200+ applicants to extend DeepChem’s support for large-scale molecular language models.
- Extended the HuggingFaceModel class by implementing the generate() API, enabling batched autoregressive inference, tokenizer padding, and device-aware execution.
- Developed an OLMo wrapper class supporting regression, classification, multitask learning, and causal language modeling with custom prediction heads, pretrained checkpoint loading, mixed-precision inference, LoRA/QLoRA, and multi-GPU support.
- Benchmarked QLoRA-adapted OLMo across 9 molecular datasets, achieving 93.1% ROC-AUC on BBBP and 92.2% on BACE with random splits, and 91.6% on ClinTox with scaffold splits.
Oct 2025 – Jan 2026
AI Associate Developer
Insurity · SpatialKey Team
Geospatial AI & ML
- Built geospatial preprocessing pipelines over 6M+ records, integrating climate and peril data via Pandas and GeoPandas.
- Trained LightGBM peril classifiers with SMOTE balancing, achieving +10% accuracy and F1 score.
- Built a U-Net++ change-detection system for before/after wildfire satellite imagery, adding shadow normalization and shadow masking to reduce false positives from terrain illumination.