Karthik Gaur

IRIS — Smartphone Object-Finding Tool for Visually Impaired Users

• Built a browser-based guidance system that takes a visually impaired user from “find my keys” to fingers on the object in ~30 seconds, no app install or wearables: voice input → phone sweep with haptic feedback → stereo audio hand guidance → confirmed contact.
• Used Gemini 2.5 Flash as a zero-shot object detector (finds anything described in natural language, no fixed classes), with MediaPipe hand tracking running client-side at ~30fps in WASM and only one frame per second hitting the server.

UA Parking Intelligence Platform

• Built a full-stack analytics platform as SGA Treasurer to give UA administration structured data on student parking for the first time. Python ETL pipeline over Qualtrics API, 1,766 anonymized responses, PII scrubbing, and precomputed metrics exported as JSON artifacts.
• Built a retrieval-augmented generation (RAG) system: embedded 2,191 free-text documents with gemini-embedding-001 (768-dim), cosine-ranked top-K retrieval per query, engineered generation prompts that feed retrieved quotes to Gemini with quote-level citations. Replaced an earlier keyword-overlap baseline. Next.js 16 dashboard with executive stats, interactive theme explorer, and chat interface.
• Findings presented to UA Parking administration in April 2026. Public demo runs on synthetic data with the same functionality.

Cervus — Local Agentic AI for Enterprise Data

• Designed and shipped a locally-hosted agentic AI tool that turns plain-English business questions into validated SQL against a production PostgreSQL database, making decades of lease and harvest data self-serve for non-technical staff. Projected to save ~1,900 analyst hours/year (~$70K across 6 stakeholders) and replace a paid Microsoft Copilot license.

• Performed exploratory analysis on fraud and ACH transaction data; modeled fraud signals across consumer/commercial split to detect fraud probability. Built data visualization dashboards for stakeholder review.

• Selected as a Google Summer of Code 2026 mentor for ML4SCI-DeepLense project (NSF-funded; Alabama, MIT, others), guiding contributors on deep learning for gravitational lensing.
• Included ‘Physics’ into base model architecture (Resnet-18), making it a Physics-informed Neural Network (PINN), improved AUC from 0.994 to 0.9953 for dark matter type classification on a NVIDIA H200 cluster.
• Built a self-supervised pipeline: pretrained a ViT masked autoencoder on unlabeled images, then fine-tuned one shared encoder for both classification (0.999 macro AUC) and super-resolution, showing cross-task transfer from a single representation.

• Built and deployed Python automation tool that eliminated Anaconda licensing costs and reduced manual work by 85% for finance.
• Engineered Python/SQL pipelines to automate reporting, reducing manual handling by 70% and accelerating access to business-critical insights. Designed a customer-focused Power BI dashboard for lease managers (saving ~160 hrs/year); ran user feedback sessions to refine metrics, IA, and flows.
• Built SSRS reports with complex SQL; managed test→prod deployments; ~$50K in annual savings.

• Launched a software-asset tracking feature across 11 production sites; defined requirements and success criteria, wrote Pytest+model_bakery unit tests, and coordinated rollout, delivering ~$63K/year in savings.
• Improved reliability and UX across 16 internal apps by resolving defects and enhancing search/navigation; contributed 10 PRs and 14 tracked issue fixes. Worked on fullstack – backend and frontend in Django.