Sai Guru Inukurthi Projects - Machine Learning & Web Development

The multi-agent pipeline routes through 4 specialized LangGraph nodes. The Risk Profiler uses a CatBoost classifier (100% accuracy) to segment users into risk tiers based on 25 driving history parameters.

High-risk quotes trigger the Conversion Predictor, heavily utilizing SMOTE to counter a massive class imbalance (22% bind rate). This node runs real-time explainability logic, generating SHAP, LIME, Anchors, and DiCE values in <200ms per quote.

Finally, the Premium Advisor leverages Groq Llama 3.3 70B to synthesize the ML outputs into a human-readable justification, routing the final decision for auto-approval or human underwriter escalation.

Built upon the bert-base-uncased architecture, this model is designed to detect semantic similarity beyond simple lexical overlap. Academic text pairs are concatenated with special [CLS] and [SEP] tokens before passing through 12 transformer blocks.

The system heavily optimizes padding strategies to fit within strict memory constraints while preserving contextual integrity. The pooled output is extracted and fed into a custom linear classification head to output a continuous probability score.

Training leveraged 569,033 sentence pairs from the SNLI dataset over 3 epochs with AdamW optimization and learning rate scheduling, achieving an F1 score of 0.868 on the test holdout set.

Built from scratch using a local Pygame engine replica, this AI utilizes NEAT (NeuroEvolution of Augmenting Topologies) to literally evolve a brain capable of flawlessly playing the Chrome Dino game.

Unlike standard neural networks where only weights are updated, NEAT evolves both the weights and the physical structure of the network via genetic algorithms. The population started with simple 12-input, 2-output nodes and dynamically grew hidden nodes to handle complex decisions like ducking under varying bird heights.

Implemented a rigid fitness function rewarding survival while heavily penalizing incorrect actions (e.g., -30 points per frame for failing to duck). The final optimized network easily surpasses a 10,000 score limit.

Bypasses standard, slow DOM color picking algorithms by rendering images into a hidden HTML5 Canvas and directly mapping the raw Uint8ClampedArray pixel buffer.

Once extracted, RGB vectors are mathematically translated into the perceptually uniform LAB color space. This prevents dark or muddy colors from mathematically dominating the palette generation compared to standard Euclidean RGB distance.

The CIE94 delta-E formula is then applied to calculate human-visual distance. A strict filtering pass strips out <0.5% pixel noise anomalies before a K-Means clustering algorithm cleanly groups the top 5 dominant swatches in under 200ms.

Fine-tuned the YOLOv8m (Medium) object detection architecture using PyTorch on a custom Roboflow dataset specifically annotated for heterogeneous web UI components (buttons, forms, navigation).

The computer vision pipeline utilizes OpenCV and PIL for pre-processing DOM screenshots, standardizing resolutions to 640x640 before passing them to the tensor core. The training process was optimized for memory constraints using a batch size of 2 across 100 epochs.

Built during a 12-hour hackathon, the model operates via an interactive Streamlit web interface with an adjustable confidence threshold (default 0.40), allowing users to upload screenshots and receive instant bounding box predictions.