P V S CHAITANYA

AI/ML

0+ yrs exp 81 · Excellent

About

Interested to explore the field of Agentic AI and apply concepts that intersect with training ML and DL models and using them in applications built around AI.

Skills & Expertise (16)

Python Intermediate

6.5/10

0.5

Years Exp

PyTorch Intermediate

6.5/10

0.5

Years Exp

scikit-learn Intermediate

6.5/10

0.5

Years Exp

Hugging Face Intermediate

6.5/10

0.5

Years Exp

Pandas Intermediate

6.0/10

0.5

Years Exp

NumPy Intermediate

6.0/10

0.5

Years Exp

OpenCV Intermediate

6.0/10

0.5

Years Exp

Git Intermediate

5.0/10

0.5

Years Exp

GitHub Intermediate

5.0/10

0.5

Years Exp

MS Excel Intermediate

4.0/10

0.5

Years Exp

MySql Intermediate

4.0/10

0.5

Years Exp

Java Beginner

3.5/10

0.5

Years Exp

Machine Learning Model Training Machine Learning Concepts Data Science & Machine Learning Deep Learning

Work Experience

ML Intern

Akcero

Jan 2026 - Jun 2026

Developed a self-supervised motion magnification framework using pretrained optical flow networks (ARFlow) to avoid reliance on synthetic datasets through flow-consistency based training while enabling test-time adaptation for real-world cases. Built a UNet-based video magnification model with positional encoding for the magnification factor (α), allowing a single model to generalize across continuous motion amplification ranges with reduced artifacts. Used optical-flow displacement fields and FFT to extract dominant vibration frequencies, amplitudes, and heatmaps for motion analysis. Built a retrieval pipeline with semantic-chunking, multi-vector embeddings, cross-encoder re-ranking with metadata generation such as {keywords, hypothetical student queries}, and query-expansion achieving sub-500ms syllabus-grounded retrieval with guardrails to prevent students from going off-topic. Built a multi-agent educational RAG system using LangGraph for persistence with supervisor-based routing and with three tutoring agents for Homework, Learning and Exam modes.

ML Intern

NRSC / ISRO

Jun 2025 - Jul 2025

Handled data preprocessing, class imbalance, and performance evaluation to derive actionable insights from model outputs. Developed a suite of 3 deep learning models (Swin Transformer, EfficientNet-B1) for automated jute crop phenotyping. Architected a multiscale context approach for identifying the faint ‘Line sowing’ patterns while achieving a 0.85 weighted F1-score on sowing classification. Also tackled severe (7.5:1) data imbalance between the classes of Broadcasting and Line sowing. Architected a novel two-stage cascaded Swin Transformer to classify crop health under extreme imbalance (2.5% ‘Poor’ class), achieving 0.64 recall for the critical minority state using Focal Loss and Oversampling. Built a multi-modal Swin Transformer by fusing image and temporal data for ordinal growth stage classification, boosting recall on the rare ‘Maturity’ class to 0.67 and achieving a 0.72 weighted F1-score.