Shriprasad
Patil

Overview

Turn research papers into visual, evidence-grounded, verifiable explanations.

Problem

Dense research papers make it difficult to trace claims back to the source material while reading, comparing, and synthesizing work.

Architecture & engineering

• Separates ingestion, parsing, provenance, extraction, verification, and rendering into explicit stages.
• Combines lexical BM25 and optional semantic retrieval with hybrid reciprocal-rank fusion.
• Supports evidence registry artifacts, paper-aware chat, workspaces, comparison, citation graphs, and a bounded research agent.

Reliability / security

Claims remain explicitly unverified when evidence or model credentials are unavailable; evaluation reports are preliminary by design.

Overview

A secure coding-agent platform where AI agents work inside isolated sandboxes, validate changes, and open approval-gated GitHub pull requests.

Problem

AI coding workflows need a controlled path from repository access through edits, validation, and human-approved publication.

Architecture & engineering

• Runs a bounded inspect → edit → validate loop through constrained tools in ephemeral Docker sandboxes.
• Connects GitHub App installations and repository permissions to ownership-aware execution jobs.
• Streams authenticated run status, tool activity, and output through WebSockets with REST recovery.

Reliability / security

Uses Argon2id password hashing, hashed session tokens, rate limiting, signature-verified webhooks, bounded logs, cancellation, and approval-gated Git operations.

Overview

An accessibility-focused AI system that transforms documents and visual content into more understandable formats through OCR, simplification, and speech.

Problem

Reading-heavy content is often inaccessible to people with dyslexia, learning differences, visual strain, or low reading confidence.

Architecture & engineering

• Processes images and PDFs through preprocessing, OCR, text cleaning, simplification, and voice generation.
• Offers adjustable reading settings, dyslexia-friendly presentation, high contrast, voice tones, and local session history.

Overview

A PyTorch image-classification system that pairs custom ResNet modeling with interpretable visual diagnostics.

Problem

Image classifiers are more useful when their predictions can be inspected, explained, tested, and served consistently.

Architecture & engineering

• Trains a custom residual network on Tiny ImageNet and exports models for ONNX-based inference.
• Includes Grad-CAM, feature-map, filter, and architecture visualizations alongside FastAPI serving and Docker packaging.