Kaviya N S

Experience

Camera-only gym analytics

Jul 2025 – Dec 2025 Chennai, Tamil Nadu Internship

AI Engineer Intern

Smartan FitTech · AI sports-tech startup turning ordinary cameras into coaching systems

Equipment recognition

Replaced frame-by-frame labelling with an automated annotation pipeline built on Segment-and-Track-Anything that exports YOLO-ready masks for dumbbells and kettlebells.
Fine-tuned YOLOv8 detection and segmentation models and evaluated them with mAP@0.5, precision, recall and IoU.
Read dumbbell weights from coloured ribbons using K-Means clustering in HSV space, which holds up under gym lighting.

3D pose estimation

Reconstructed SMPL body meshes (6,890 vertices) from exercise video and classified form as correct or incorrect with a Bi-LSTM.
Pruned non-joint mesh regions around 17 YOLOv8-Pose keypoints to balance inference latency against accuracy.
Added an OAK-D depth camera to track joint drift that 2D video misses, and compared accuracy with and without depth input.

Selected projects

Reported like model cards

For each project: the problem, how I approached it, the evidence that it works, and what it runs on.

AskPDF

LLM · retrieval-augmented generation

code ↗

Problem: Ask questions across several PDFs, including scanned and two-column papers, without the model making things up.
Approach: Semantic chunking, Pinecone retrieval and Llama 3.1, with intent routing between Q&A and summaries plus OCR fallback.
Evidence: NLIfaithfulness score on every answer, plus coverage and abstractiveness

LlamaIndexPineconeGroqTransformersStreamlit

CipherTrade

Deep reinforcement learning

code ↗

Problem: Train an agent to trade BTC/USD on its own, from what a chart looks like.
Approach: Soft Actor-Critic with a CNN encoder over 128×128 candlestick images, twin critics and automatic entropy tuning, in a custom Gymnasium environment.
Evidence: MLflowreward, P&L, losses and checkpoints logged for every episode

TensorFlowGymnasiumMLflowStreamlit

Deep-Shield

Computer vision · explainable AI

code ↗

Problem: Detect AI-generated images in a way that can be explained and is hard to fool.
Approach: Fine-tuned ResNet-18 on CIFAKE, Grad-CAM explanations, FGSM, blur and frequency attacks, then adversarial training.
Evidence: 0.81accuracy · 0.93 recall on fakes (2,000 held-out images)

PyTorchGrad-CAMscikit-learn

Pulse

Information retrieval · recommenders

code ↗

Problem: Rank social posts by quality instead of popularity.
Approach: Boolean, TF-IDF and BM25 search, synonym query expansion, a quality classifier, KMeans clustering and two recommenders.
Evidence: 0.61 → 0.99nDCG with query expansion · P@5 0.92 vs 0.17 random

Djangoscikit-learnNLTKrank-bm25

RoadSoS

Full-stack · geospatial

code ↗

Problem: Help arrives too late after road accidents, often because nobody can call.
Approach: A phone web app that detects crashes from GPS speed and accelerometer spikes, then texts family with a location link.
Evidence: 15,000+hospitals, police and rescue services searchable across Tamil Nadu

ReactFlaskPostGISLeafletTwilio

Outpost

Backend · API design

code ↗

Problem: Publish one piece of content to many niche platforms through a single API.
Approach: FastAPI connector contract, rules engine and job runner, packaged with Docker Compose and a certification test suite.
Evidence: 6destinations: Discourse, Lemmy, Dev.to, Ghost, webhooks, manual packs

FastAPIPostgreSQLDockerpytest

More work	What it is	Result / stack
Deep Isolation Forest	Paper implementation for unsupervised credit-card fraud detection	ROC-AUC 0.94
SpokenAI	Speech coach scoring relevance, coherence, fluency and vocal emotion	Whisper · SBERT
InvoiceSight	Field and line-item extraction from scanned invoices with a verifiability report	PaddleOCR
Face Recognition + RAG	Live webcam recognition with a chat that answers questions about registrations	LangChain · FAISS
Fleet Manager	Multi-robot traffic control with Dijkstra routing and lane locking	Python · Tkinter
RAPID	Flood-response command platform with GIS risk zones for Chennai	React · PostGIS

Toolkit

What I work with

Languages

Python, SQL, Java, JavaScript, C++

Machine learning

PyTorch, TensorFlow / Keras, scikit-learn, NumPy, pandas, SciPy

Computer vision

YOLOv8, OpenCV, SMPL, Grad-CAM, segmentation & tracking

LLMs & NLP

LlamaIndex, LangChain, Pinecone, FAISS, Hugging Face, Whisper

MLOps & backend

MLflow, FastAPI, Flask, Django, Docker, pytest, Git

Data

PostgreSQL / PostGIS, MySQL, MongoDB, ClickHouse

Education

Training

Aug 2022 – present Integrated M.Sc. Data Science (5-year) PSG College of Technology, Coimbatore

Certification Machine Learning Specialization DeepLearning.AI · Coursera

Certification Natural Language Processing Specialization DeepLearning.AI · Coursera

Hackathons

Built under deadline

GOAT Robotics × PSGMulti-robot fleet management system

KatomaranFace recognition platform with RAG chat

IncuBrixOutpost, a long-tail content distribution API

Yavar.aiInvoiceSight, invoice OCR and verification

Contact

Let's talk about your ML team

I'm looking for full-time roles in data science, machine learning, computer vision or AI engineering, starting December 2026.

kaviya.ns.tech@gmail.com GitHub ↗