Skip to content
View Prishita01's full-sized avatar
🎯
Focusing
🎯
Focusing
6 followers · 19 following

Highlights

  • Pro

Block or report Prishita01

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Prishita01/README.md

Prishita Ghanathe Krishna

MS Computer Science (AI/ML), Case Western Reserve University. I build ML systems end-to-end — inference pipelines, orchestration, monitoring, retraining. Not just models.

Projects

perception-av — Production AV perception pipeline on nuScenes. YOLO26 detection (2,221 detections, 104ms/frame), Depth Anything V2 depth maps in S3, SAM2 pixel masks in PostgreSQL. Airflow 2.9 across 7 Docker containers, Evidently AI drift monitoring, hard case mining, MLflow tracking, YOLO26n fine-tuning (+251% mAP50 on held-out val), dbt + Grafana analytics layer.

arxiv-rag — Production RAG over arXiv papers. Hybrid retrieval: BM25 sparse + sentence-transformer dense, fused with RRF. Cross-encoder reranking, Llama 3.2 via Ollama, FastAPI, Docker Compose.

ReconStructAI3D — Monocular 3D reconstruction from smartphone photos. SAM segmentation, COLMAP SfM, OpenMVS dense reconstruction, Open3D refinement. PSNR 24.3 dB, SSIM 0.985, no depth sensors.

Triton-ResNet50 — Three custom Triton GPU convolution kernels benchmarked across 50+ models on NVIDIA H100 NVL and L40S. ScalarMM outperforms cuDNN ~29% on ResNet-family layers. Validated at ~1e-5 numerical correctness end-to-end.

FairCredit — Responsible AI framework for credit risk. 68.7% bias reduction at 2.3% accuracy cost across three datasets. SHAP explainability, post-processing threshold optimization.

StockNet HPC — Distributed LSTM training with PyTorch DDP + SLURM. 72.6x CPU-to-GPU speedup (5515s to 76s) on CWRU Markov HPC cluster.

Stack

Python, PyTorch, Apache Airflow, Docker, AWS S3/RDS, YOLO26, SAM2, Depth Anything V2, MLflow, Evidently AI, dbt, Grafana, FastAPI, PostgreSQL, OpenAI Triton, CUDA, COLMAP, OpenMVS, SLURM, Hugging Face

Research

Prishita G K. Eye Tracking: Gaze Detection in Market Research. IJIRCCE Vol. 11(5), 2023.
Prishita G K. AutoML: Automated Machine Learning using Streamlit. IJIRCCE Vol. 11(5), 2023.

prishita.ghanathe@gmail.com · linkedin.com/in/prishitagk
Research Assistant @ CWRU KLab-AI3 · open to ML Engineer / CV Engineer / MLOps roles

Pinned Loading

  1. perception-av perception-av Public

    AV sensor data pipeline involving Airflow orchestration, YOLO26/SAM2/Depth inference, drift monitoring, and MLflow-tracked retraining on nuScenes dataset.

    Python 1

  2. arxiv-rag arxiv-rag Public

    Production RAG system for ArXiv papers with hybrid retrieval, cross-encoder reranking, and local LLM generation

    HTML 1

  3. Triton-ResNet50 Triton-ResNet50 Public

    Python 1

  4. ReconStructAI3D ReconStructAI3D Public

    Monocular 3D reconstruction pipeline using SAM + COLMAP + OpenMVS — PSNR 24.3 dB, SSIM 0.985, no depth sensors.

    Python 3 1

  5. Stock-Prediction-GPU Stock-Prediction-GPU Public

    Distributed multi-GPU LSTM/CNN training pipeline on HPC. 72.6× CPU→GPU speedup with PyTorch DDP + SLURM.

    Python

  6. FairCredit FairCredit Public

    Responsible AI framework for credit risk. 68.7% fairness improvement at 2.3% accuracy cost. Validated across 3 datasets

    Jupyter Notebook