Skip to content

Projects

Featured project

Efficient Test-Time Scaling through Execution-Supervised Refinement

Sep 2025

Gemini APILLMsExecution-Supervised Refinement
  • Designed Self-Refine-Loop, a Gemini 2.0 Flash pipeline that generates, executes, and repairs outputs using deterministic feedback from a secure Python execution environment, with up to three refinement iterations.
  • Achieved 88.3% MBPP, 98.3% GSM8K, and 76.7% BoolQ accuracy with 2.5 model calls per problem, outperforming MiPS-style best-of-32 selection with over 90% fewer calls.
  • Implemented Dynamic Temperature Adjustment and Context Injection to adapt to failure types and supply missing helper classes, raising MBPP accuracy from 40.0% to 88.3%.
View code Read paper (PDF)

Research details

Efficient Test-Time Scaling through Execution-Supervised Refinement

Question
Can execution feedback help an LLM repair a solution with fewer model calls than generating many candidates?
Method
Generate an answer with Gemini 2.0 Flash, evaluate it, and feed failure signals into up to three refinement iterations. Context injection supplies helper definitions, while temperature adapts to the error type.
Reported MBPP accuracy: Self-Refine 88.3%, MiPS reference 57.8%. GSM8K: Self-Refine 98.3%, MiPS reference 80%.
Coding and math accuracy compared with the MiPS results cited in the paper. Select the image to view it at full size.

Evaluation: 60 MBPP, 60 GSM8K, and 30 BoolQ problems. MiPS values are literature references using different model and evaluation setups.

Mid-Day Meal Analysis using Machine Learning

Jan 2023

KNNLogistic RegressionLightGBM
  • Analyzed nutrition and program data across 24 Indian states; KNN and Logistic Regression models achieved 94.2% and 91.6% accuracy for nutrition-related prediction and child meal-outcome analysis.
  • Trained LightGBM classifiers for rice, wheat, and dry-bean quality inspection, achieving 99.88%, 92.52%, and 93.22% accuracy; serialized models with Pickle for web-app integration.
Read paper Local PDF

Research details

Mid-Day Meal Analysis using Machine Learning

Question
How can nutrition requirements, meal outcomes, and ingredient quality inform India’s Mid-Day Meal program?
Method
Use K-Nearest Neighbors to predict protein requirements and Logistic Regression to evaluate meal outcomes. Separate LightGBM classifiers inspect rice, wheat, and dry-bean quality.
Findings
The study reports 94.2% accuracy for KNN and 91.6% for Logistic Regression. Food-quality classifiers achieved 99.88% for rice, 92.52% for wheat, and 93.22% for dry beans.
Application
Models were serialized with Pickle for integration into an interactive web application.

Alzheimer’s Disease Detection & Categorization

Aug 2022

Logistic RegressionDecision TreesDjango
  • Compared Logistic Regression and Decision Tree classifiers for dementia categorization, reporting 95% and 91% accuracy, respectively, in the study.
  • Analyzed feature importance and PCA projections, and integrated the Logistic Regression model into a Django application using Pickle.
Read paper Local PDF

Research details

Alzheimer’s Disease Detection & Categorization

Question
How can clinical assessment features support the classification of dementia-related patient data?
Method
Compare Logistic Regression with a Decision Tree tuned using Grid Search CV. Inspect feature importance and PCA projections, then integrate the prediction model into a Django form.
Findings
The paper reports 95% accuracy for Logistic Regression and 91% for the tuned Decision Tree. CDR and MMSE were among the important assessment features.
Application
The Alzem-cure research prototype combines model predictions with pages containing games and activities. These are study results, not clinical validation.

Brain Tumor Classification Pipeline

Dec 2021

KNNPCAPlotly Dash
  • Tuned a K-Nearest Neighbors classifier on a 699-row, 11-column tumor dataset, evaluating K values from 1–50 with cross-validation and achieving 97% classification accuracy.
  • Applied PCA to reduce features into two- and three-dimensional spaces; used Plotly Dash to compare actual and predicted class separation and investigate misclassifications.
Read paper Local PDF

Research details

Brain Tumor Classification Pipeline

Question
How well can a KNN classifier distinguish tumor classes, and what can visualization reveal about its errors?
Method
Evaluate neighborhood sizes from 1–50 with cross-validation, then use PCA to project features into two and three dimensions for comparison of actual and predicted classes.
Findings
The paper reports 97% accuracy after tuning, with four false positives and one false negative in its test confusion matrix.
Visualization
Interactive Plotly Dash views expose class separation and misclassified samples in the research dataset.

Automatic Rain-Shade Simulation (Arduino + Proteus)

Jun 2020

EmbeddedSimulationAutomation
  • Simulated an UNO + rain sensor controlling dual servos for a retractable roof.
  • Modeled the roof mechanism in AutoCAD 2020 and simulated rain-triggered 90-degree servo movement in Proteus.
  • Documented circuit, algorithm, and test cases for home-automation use.
Read paper (PDF) Conference record

Research details

Automatic Rain-Shade Simulation (Arduino + Proteus)

Question
Can an automatic roof protect clothes on a clothesline when rain begins while the user is away?
Method
Simulate an Arduino UNO, a rain sensor, and two servo motors in Proteus. Model the mechanical arrangement in AutoCAD 2020.
Operation
A rain signal triggers a 90-degree rotation of the servo motors to move the shade. When the sensor returns to its dry state, the servos return to their original position.
Outcome
The paper documents the circuit simulation and mechanical design for a home-automation concept; it does not report a field-tested hardware deployment.