Skip to content
View Parth310800's full-sized avatar

Block or report Parth310800

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
Parth310800/README.md

Parth Abhishek Mehra

Data Scientist — forecasting, NLP, and the evaluation rigor that makes model results trustworthy

M.S. Data Science, Indiana University · Atlanta, GA · Open to relocation

Email LinkedIn


Most models fail quietly. They report a strong number while doing something useless, and the number is convincing enough that nobody checks.

I spend a lot of my time on the checking. Each project below began as a straightforward modelling exercise and became something more interesting once the evaluation was examined properly.


A classifier reporting 97% accuracy — while missing 94 of 181 actual frauds.

At 4.8% prevalence, predicting "legitimate" for everything already scores 95%. Rebuilt the pipeline to eliminate train/test leakage in the vectoriser, report PR-AUC against a majority baseline, and select the decision threshold from the precision–recall curve rather than defaulting to 0.5.


Three prompting strategies for persona imitation, benchmarked on a 3.8B model under 4-bit quantization on a single T4.

Retrieval's advantage on semantic alignment held under paired testing. The stylistic-fidelity result did not — and the metric measuring it turned out to track response length rather than persona markers. The measurement problem became the finding.


Synthetic minority-class generation for a fraud dataset at 0.13% prevalence.

Standard TSTR evaluation with ROC-AUC suggested the synthetic data worked. Evaluating at true prevalence with PR-AUC showed otherwise. Distance-to-closest-record analysis put synthetic samples ~27× further from real records than real records sit from each other — strong privacy, at a measurable utility cost.



🛠 Toolkit

Python SQL R

scikit-learn TensorFlow PyTorch Hugging Face

pandas NumPy Spark Airflow

AWS Docker Power BI


📍 Currently

Benchmarking transformer architectures for clinical concept extraction on the NBME patient notes dataset — TF-IDF, BERT-base, and DeBERTa-base for token-level span extraction, with grouped splits by patient note to avoid leakage across repeated entities.

Background

M.S. Data Science, Indiana University — 2026 B.Tech, Vellore Institute of Technology

Roughly 2.5 years in industry across data science, data engineering, and automation before graduate school, including demand forecasting work in manufacturing operations.


Open to Data Scientist, Machine Learning, Data Analyst and Data Engineer roles.

If you're hiring — or just want to talk about why your model's metrics might be lying to you — email me.

Pinned Loading

  1. linguistic-fingerprinting linguistic-fingerprinting Public

    Persona imitation benchmark with a quantized local LLM — and why standard style-fidelity metrics don't work

    Jupyter Notebook

  2. nbme-clinical-concept-extraction nbme-clinical-concept-extraction Public

    Comparative study of TF-IDF, BERT, and DeBERTa for character-level clinical span extraction, with leakage-corrected evaluation and confidence-based human review triage.

    Jupyter Notebook