Machine learning and analytics projects
Every project is evaluated on the metric that genuinely reflects its objective: F1-Macro for severe class imbalance, MAP@3 for candidate ranking, and validation F1 for customer retention. High accuracy on lopsided data can flatter a model, so I lead with the score that shows how it does on the rare cases.
Featured projects
-
Smart MCQ solver (LoRA fine-tuning)
Parameter-efficient fine-tuning of DeBERTa-v3 and RoBERTa using LoRA adapters for 5-option multiple-choice ranking. Blended in a 70:30 ensemble achieving 0.74242 MAP@3 on 500 unseen questions.
0.74242 MAP@3 -
Comment classification
Feature engineering over 198,000 comments across four imbalanced label classes. Engineered 10,000+ n-grams, screened tabular features with chi-square and Kruskal-Wallis, and tuned LightGBM with Optuna (0.8202), then blended 3 models to reach 0.8293 F1-Macro.
0.8293 F1-MACRO -
Customer exit prediction
I compared 8 classifiers on 90,000 customer records from a financial institution, where few customers leave. I filled missing values, then tuned a Random Forest to a validation F1 of 0.6365 (accuracy 0.8414).
0.6365 F1 -
VC deal analytics dashboard
Power BI deal analytics dashboard built on a 50,000-row synthetic star schema. Engineered 7 DAX measures and configured 2 AI visuals (Key Influencers and Decomposition Tree) for valuation driver ranking.
POWER BI · STAR SCHEMA
More projects, in brief
-
Virtual AI Teaching Assistant (RAG Pipeline)
Built a Retrieval-Augmented Generation (RAG) assistant using OpenAI embeddings and SQLite vector store. I split course materials into chunks and retrieved top-k relevant context via cosine similarity with FastAPI endpoints.
2025 · RAG · FASTAPI -
HR Attrition Dashboard (Power BI)
A two-page Power BI dashboard with 12 DAX measures on staff turnover, comparing hires with the same quarter before.
2026 · POWER BI · DAX -
Consumer Spend Segmentation (K-means++)
I grouped 62 survey respondents into 5 segments by monthly spend on personal care and fast food, using K-means++ (silhouette 0.540). Adding more spend categories and age made the segments less distinct, so I kept two.
2024 · K-MEANS++ · PYTHON -
Migration Determinants Regression Study (R)
Specified a 5-regressor OLS model in R across 28 Indian states analyzing out-migration. Found per-capita income (p=0.002), poverty (p=0.025), and unemployment (p=0.018) statistically significant at 5%.
2025 · R · OLS REGRESSION -
Influencer and Sponsor Matching Platform (Flask)
Full-stack web application connecting content creators and commercial sponsors using Flask, SQLite, and Jinja2 with 10+ analytical tracking dashboards and role-based access controls.
2024 · FLASK · FULL-STACK