Kairav Mittal is an AI Engineer and Software Engineer specializing in Large Language Model alignment, bias mitigation, and explainable computer vision systems. He builds production-grade AI solutions including DPO-fine-tuned LLMs and hybrid deepfake detectors. Experienced open-source contributor and AI quality assurance specialist with a proven track record of shipping impactful, ethical ML projects that outperform baselines and deliver clear explainability.
Engineered a complete debiasing pipeline for facebook/opt-1.3b using Counterfactual Data Augmentation + Direct Preference Optimization (DPO) with 4-bit QLoRA. Reduced stereotypical preference on CrowS-Pairs benchmark by 3.7 points. Released fine-tuned model on Hugging Face + live Gradio demo.
Fine-tuned google/vit-base-patch16 on 160k real/fake images for robust AI-generated face detection. Hybrid system fusing ViT predictions with classical forensics (ELA, noise mapping, LBP texture) plus Grad-CAM heatmaps.
Comprehensive desktop application for time series experimentation, anomaly detection, and forecasting. Provides an out-of-the-box environment with multiple analysis pipelines, interactive visualizations, and Dockerized deployment.