Build it. Test it. Trust it.

I build machine learning systems, then spend just as much time trying to break them.

Data Science & AI at IIT Madras and Computer Science Engineering at GNDEC. I take models from notebook to cloud, with particular attention to data quality, evaluation, and whether a result actually deserves to win.

See live work Download resume

selected work

Two systems you can open.

These are working deployments, not only repository descriptions. One ranks answers; the other serves and compares trajectory models through an interactive dashboard.

ranking system Smart MCQ Solver interface showing a RoBERTa Base model and question-entry workflow

Smart MCQ Solver

A ranking system optimized for MAP@3, taken through six milestones from TF-IDF baselines to RAG and LoRA fine-tuning. The strongest model was a LoRA fine-tuned RoBERTa-base, deployed as a Streamlit app.

stackPython · PyTorch · Transformers · PEFT/LoRA · Streamlit

model RoBERTa-basemetric MAP@3deploy Streamlit

0.75311 MAP@3 3.4 MB adapter

The useful result was not just the score: DeBERTa-v3, the apparently stronger pretrained model, failed to converge with its adapter at about 17% train accuracy.

Open live demo GitHub

sensor fusion INS and GPS Trajectory Fusion dashboard showing loaded models and trajectory comparison charts

INS/GPS Trajectory Fusion

A full-stack system for predicting autonomous-vehicle trajectories from IMU and GPS input. Twelve architectures are served through FastAPI with CPU-optimized PyTorch weights, all under 6 MB.

stackPython · PyTorch · FastAPI · NumPy · Pandas

models 12 architecturesmetric displacement errordeploy FastAPI

12 architectures 11 evaluation metrics 5.58 m average displacement error

The dashboard supports synthetic path simulation and CSV upload, making model comparison something a visitor can inspect rather than take on faith.

Open live demo GitHub

Unified Intelligent System for Audio Inference

stack / Python · FastAPI · LangChain · Whisper · YAMNet

A modular pipeline combining Whisper large-v3 for speech transcription and YAMNet across 521 AudioSet classes. Multi-stage filtering reduces 66 raw detections to 8–10 clean segments before LLM reasoning.

Multi-Crop Disease Detection System

stack / Python · TensorFlow · Keras · EfficientNetB0

Classification across 8 crops grown in Northern India and 37 disease classes. Transfer learning with EfficientNetB0 reached 0.94 test accuracy, compared with a 0.54 F1 CNN-from-scratch baseline.

Sentiment Classification & Error Analysis

stack / Python · Scikit-learn · LightGBM · XGBoost

A reusable evaluation workflow using Logistic Regression, Scikit-learn, LightGBM, and XGBoost, with feature engineering and tuning improving F1 by 0.15 over baseline and tracing errors to labeling inconsistencies.

education + experience

Where the work got built.

Education and applied research in one timeline: the foundations at IIT Madras and GNDEC, followed by internships where those foundations became working systems.

2023–2027
GNDEC / CSE

Bachelor of Technology

Computer Science Engineering · GNDEC

Computer Science Engineering at GNDEC.

2023–2026
IIT Madras

Diploma in Data Science and Its Application

IIT Madras

Data Science and its application at IIT Madras.

Jun–Jul 2026

CSIR-CMERI, Ludhiana

AI/ML Research Intern

Built IMU+GPS sensor-fusion pipelines for trajectory and navigation state estimation. Compared 12 architectures under a leakage-aware protocol across 11 metrics; BiMamba reached 5.58 m average displacement error.

Jun–Jul 2025

Punjab AI Excellence, Ludhiana

Data Science Intern

Under Dr. Sandeep Singh Sandha, fine-tuned EfficientNetV2B0 for 37-class plant disease classification. Audited label noise and class imbalance; cleaned-label retraining cut misclassification by about 12% on held-out validation.

skills

Tools behind the work.

Languages

Python, SQL, C++

ML / deep learning

Scikit-learn, TensorFlow, PyTorch, Pandas, NumPy, Matplotlib

Systems

FastAPI, LangChain, librosa, LoRA/PEFT

Tools & cloud

Git/GitHub, Jupyter, Linux, Google Colab, FFmpeg, AWS

resume

A concise record of the work.

Review my experience, education, projects, and technical skills in one place.

Download resume

contact

Let’s talk about the work.

Open to entry-level Machine Learning Engineer, Data Scientist, Data Engineer, and AI Engineer roles — remote or based in Bangalore, Gurgaon, or NCR.

aryanpraja004@gmail.com LinkedIn GitHub