SAIANIRUTH M

Binary classification system for pediatric chest X-rays trained on ~60K annotated images (20K positive, 40K negative). Two-tower feature fusion of DenseNet-169 (1024-d) and EfficientNet-B2 (1408-d) into a 2432-d concatenated representation, followed by a two-layer classifier head (Linear→ReLU→Dropout(0.3)→Linear→Sigmoid). Preprocessing pipeline applies LANCZOS resizing to 224×224 and CLAHE contrast enhancement. Trained with BCELoss + Adam (lr 1e-3) over 50 epochs on an 80/10/10 split with autobatch sizing, comparing standalone DenseNet/EfficientNet against the fused dual-backbone variant.

Fracture detection system using YOLOv8, YOLOv11, and YOLOv12 architectures. Evaluated both standard detection (DET) and Oriented Bounding Box (OBB) models for rotated fractures.

Object detection using EfficientDet family (D0-D7x) with PyTorch Lightning. Compared model variants based on backbone size and dataset balancing strategies.

Multiclass classification (Clavicle, Humerus, Scapula, Other) using Detectron2 (Faster R-CNN) and RF-DETR Large. Handled class imbalance with hard negative mining.

Multi-architecture deep-learning framework for automated shoulder fracture localization on AP radiographs, subject of an arXiv publication. Integrates three independently-trained detectors — Faster R-CNN (ResNet-50 FPN baseline for high-quality localization), EfficientDet-D3 (BiFPN + EfficientNet backbone for fast inference), and RF-DETR (pure-transformer head with global attention for maximum recall on subtle cortical disruptions) — and fuses their outputs through an IoU-clustered, confidence-weighted ensemble that anchors final geometry to RF-DETR boxes. Preprocessing covers intensity normalization, histogram equalization, automated shoulder-joint cropping, and resolution standardization. Evaluation includes mAP at multiple IoU thresholds, ROC/PR curves, and dedicated false-negative analysis for hairline, greater-tuberosity, and subtle cortical-contour fractures. Production-ready: ONNX/TensorRT export with ensemble inference at 120–250 ms on A100-class GPUs.

Comprehensive evaluation of multiple architectures (SIGLIP, ResNet, DenseNet, EfficientNet) with fusion methods and Decision Rules Engine for general fracture detection.

Classification system for Tuberculosis (Active vs. Chronic) and Silicosis. Orchestrated 12+ pathology classifiers with complex logic rules for final diagnosis.

Critical care pathology detection for Pneumothorax, Lung Collapse, and Subcutaneous Emphysema. Compared segmentation (YOLO-Seg) vs detection (RF-DETR) approaches.

Specialized object detection models for Mass, Fibrosis, Rib Fracture, Pneumothorax, and Mediastinal Shift. Dedicated models for high-priority pathologies to maximize sensitivity.

Large-scale training on 120K chest X-ray images. Normal vs Abnormal classification and segmentation (Pleural Effusion, Consolidation) using diverse backbones.

Production-ready multi-label classification system for 40 pathologies using YOLO11-M backbone with FPN. Designed with a configurable system allowing seamless switching between Weighted BCE and Asymmetric Focal Loss via YAML to handle class imbalance. Features two-phase fine-tuning, mixed precision training, and ONNX/TensorRT export.

Quality-gate pipeline for detecting suboptimal chest X-rays before they enter the diagnostic queue. Combines a fleet of multiclass quality classifiers with OR-aggregation logic for higher failure sensitivity, plus a Gemini-based structured QC layer that consumes DICOM metadata and PNG renders to extract exposure, artifacts, cropping, rotation, flipping, mobile-capture, age-group, and overall-quality verdicts under strict JSON schemas. Includes deterministic field-validation and QC-label mapping for downstream parsing reliability.

Performance analysis of Google's CXR Foundation model (pre-trained ViT) for medical imaging embeddings. Evaluated transfer learning capabilities and bottlenecks.

Technical evaluation of Stanford AIMI's CheXagent-8b for zero-shot binary disease classification (pneumothorax, pleural effusion). Multimodal vision-language model assessment.

Fine-tuned multimodal LLMs (PaliGemma 3B, Gemma 4B) for automated radiology report generation from chest X-rays. JSONL datasets with image-text pairs. Architecture selection informed by extensive Vision Encoder experiments evaluating XrayGPT and BioMedLM for efficiency.

Fine-tuned PaliGemma (3B, 10B, 28B) for shoulder fracture detection and bounding box localization. QLoRA for efficient training with custom attention heads.

Contrastive learning framework based on CLIP to align medical images with diagnostic text. Trained on 420K samples across 69 labels for multi-label classification.

Comprehensive knee X-ray analysis system generating structured medical reports. Modular pipeline with bone detection, view classification, and pathology detection.

End-to-end productionized CXR diagnostic system, evolved from a single inference backend into a four-component product (CXR inference, orchestrator, React frontend, packaging) shipped as both a Windows desktop installer and a Linux web deployment. Modular DB-driven inference replaced the original flat-file monolith with four packages (pathologies, malposition, reporting, security), RAM-adaptive model management, hot-reload of thresholds via DB watcher, persistent SQLite-backed CXR queue with stuck-process detection, SSE streaming with a /health state machine, rotation-aware tiered CTR clinical decision logic, AES-256 streaming model decryption, and a per-detection feedback + analytics layer with cloud sync. Windows track ships via Electron + PyInstaller frozen builds, install-time model decryption, ProgramData-anchored logging, network precheck wizard, and an admin-aware auto-updater with tiered mandatory-grace policy. Linux track runs behind nginx + certbot, systemd-supervised services, Firebase auth with email-domain allowlist, and an aggregated health endpoint.

RAG-style system using MetaCLIP for retrieving relevant historical X-ray reports. FAISS-indexed 1.6M images with Gemini 2.5 Flash for report synthesis.

Comprehensive ICU-focused chest X-ray analysis system detecting critical pathologies and device malpositions. Includes live inference backend with FastAPI and interactive 3D web showcase for conference demonstration.

Comprehensive AI solution for Chest X-Ray analysis including pathology detection, device detection, malposition analysis, and cardiothoracic ratio.

Standalone FastAPI services for chest X-ray findings. Independent deployable apps for various pathologies with standardized API contracts.

Production-grade API analyzing chest X-rays for supportive devices and malposition analysis with real-time SSE streaming.

A full deep-learning pipeline for detecting scoliosis from spinal X-ray images. Multiple CNN architectures implemented (ResNet/DenseNet/ViT variants). Includes Grad-CAM for interpretability at block/feature-map level. Preprocessing: resizing, noise removal, intensity normalization. Pediatric vs Non-Pediatric classification capability. Evaluated using accuracy, recall, precision, confusion matrix. Designed to scale to 250K medical CXR images.

Generates large-scale synthetic ID card images for Computer Vision model training. Template-based ID generation. Randomized text, dates, names, photos for maximum variety. Supports exporting datasets for OCR/fraud detection. Modular design for adding new templates quickly.

A clustering solution to categorize customers using behavioral attributes. Uses Annual Income + Spending Score. Elbow Method used for optimal cluster count. Visualizations for cluster boundaries. Helps identify high-value vs low-value customer groups.

Predicts stock closing prices using an LSTM neural network. Normalization using MinMaxScaler. Time-window sequence generation. Multi-layer LSTM with dropout. Visualized predicted vs real closing prices.

A regression machine learning project focused on price prediction using the CarPrice_Assignment.csv dataset. Predicts car prices using features such as engine size, horsepower, curb weight, and fuel type. Builds multiple models: Linear Regression, Decision Trees, Random Forest. Includes EDA, correlation analysis, and outlier treatment. Measures performance using MAE, RMSE, R².

Data engineering and visualization across five years of global happiness data. Unified 2015–2019 datasets. Standardized inconsistent features (e.g., 'Economy (GDP per Capita)' → 'GDP per capita'). Visualizes GDP, Trust, Social Support vs Happiness Score. Heatmaps reveal strongest global indicators.

Analysis of a multi-period sales dataset using classical time-series techniques. Decomposition into trend, seasonality. Tested ARIMA/SARIMA models. Rolling means + stationarity tests (ADF).

Interactive business analytics dashboard. Sales performance metrics. Customer segmentation. Product profitability and trends.

Credit card transaction analysis using SQL. Window functions, joins, aggregates. High-spend customers vs category analysis. Potential fraud detection patterns.

A secure local credential management tool. AES encryption. Local encrypted vault storage. GUI + CLI interfaces. Auto password generator. Integrity check + safe read/write.

End-to-end AI-driven ATS with multi-stage recruitment screening. Gemini-based resume scoring with dynamic MCQ generation tailored to candidate profiles.

A full pipeline for scraping Amazon/Flipkart/Myntra. Handles lazy loading, pagination. Rotating proxies to avoid blocks. CAPTCHA bypass support. Exports structured JSON. GPT-powered product comparison, shortlisting, summaries.

Automates multi-annotator validation in Label Studio. Aggregates annotations. Detects annotation conflicts. Flags mismatches for review. Prepares cleaned dataset for ML training.

CLI application to manage student records. Linked list backend. File persistence. Password-protected admin interface. Console colors + formatting.

Banking simulation with role-based access. Admin + User logins. Binary file account storage. Fund transfers, balance inquiry. OOP modeling (classes, inheritance).

Early Electrical Engineering projects (Sound Activated System & Street Light Automation) representing my initial background before transitioning to AI.

Responsive personal website showcasing skills & projects.