AIE Code Agents Hackathon

CodeRalph is a real-time collaborative code editor where humans and AI agents work together as a team. Unlike traditional AI coding assistants that focus on solo autocomplete, we built a multi-agent system where Mind Reader tracks what developers are building, Shadow Test runs edge cases in the background, and Conflict Guard predicts merge conflicts before they happen. The platform is designed for scenarios where multiple people need to code together, like technical interviews where two candidates solve different parts of the same problem. Our agents don't just assist individual developers; they understand the relationships between everyone's code and help the team move fast without breaking things. It's multiplayer coding meets multi-agent AI collaboration.

Self-driving game development agency that autonomously generates new game prototypes. Just set a direction, seed code, and press start, and the agent makes games. A developer can get the leverage of an entire game dev agency even when sleeping.

BugPoint is an LLM-native debugging layer that shows AI how code actually executed instead of drowning it in logs and it integrates with your workflow. It uses program analysis with runtime instrumentation to give precise, structured visibility into real execution paths. Key capabilities include: Control-flow graphs, clear, abstract maps of program logic so agents understand structure, not just crash point Runtime traces, capture intermediate state, variable values, and basic-block paths to show where execution drifted from intent Semantic failure detection, pinpoints the exact logic break, replacing guesswork with actionable insight Live visualization, real-time views of evolving execution flow with failure nodes highlighted instantly MCP integration, lets Cursor, Claude Code, Windsurf, Zed, and other agentic tools invoke the debugger natively Minimal failure slices, compress megabytes of logs into clean, structured, LLM-ready summaries that shrink context and boost fix accuracy

Argus: A Clone brand format & re-imaginer of any website, in seconds. - Multi-Agent Code Review & Collaboration Platform - Groq fast Auto-fix - MongoDB for memory and context - Firecrawl for web extraction - Built around Google Gemini Pro 3 & Nano Banana Pro - Browser extension for advanced debugging - E2B for hosting and sandboxes - Brave for web search

Software engineering has a "Context Problem." We treat developers like routers, constantly interrupting them to ask, "Hey, how did you fix that bug yesterday?" or "Where are the API keys?" Every question destroys 20 minutes of Flow State. DevScope is an agent that eliminates these interruptions. It acts as the "Visual Cortex" for the team. It runs silently on your machine, using Gemini 2.0 Vision to observe your workflow. But it doesn't just record pixels; it extracts hard technical data like specific error codes, configuration values, and documentation titles, and syncs them to a searchable "Hive Mind" on MongoDB Atlas. The result? Async Omniscience. A new teammate can ask the DevScope Oracle, "How did Eve fix the streaming lag?" and get the exact configuration change she made 4 hours ago, without ever tapping her on the shoulder. It turns ephemeral deep work into permanent organizational memory.

An IDE extension & MCP that provides intelligent, context-aware code assistance by tracking your cursor position using AST & Gemini to pass relevant files to your AI coding assistants to help resolve issues such as getting lost in the middle and imprecise prompting from the user.

TestPlanna analyzes your code, generates test scenarios, and automates execution—all inside Jira.

Existing tools only find syntax matches we use Gemini embeddings for semantic similarity to find functionally equivalent code. We cluster similar functions with graph theory and use AI agents that generate refactoring plans with code changes and risk assessment turning days of manual work into minutes

A multi-agent AI collaboration for automated pull request reviews. When a developer opens a PR, four specialized AI agents analyze different aspects and conduct a collaborative "stand-up" discussion to provide comprehensive feedback.

Legacy codebases and new codebases are difficult to navigate with. There are some people that are good at this ability (The "Fixer" Archetype SWEs) that can make changes in surgical manner without breaking the existing solution. But for most of the others, having a knowledge graph about the codebase help them update files while knowing which files depend on it and on what files the current file depends upon. So, this helps in working around unfamiliar codebase. This also helps in on-boarding a new codebase Eventually, this idea can be scaled for microservices and we can know how all the microservices interact and we can discover optimizations like redundant calls or better load distribution

The AI Agent Mind Map Visualizer is a real-time transparency and Abstraction tool for agentic IDEs. It combines LangGraph state management with a modern React dashboard to provide interactive visualization of agent decision-making, complete with time-travel capabilities and branch management.

Braless allows for creative collaborative live coding music with Strudel.

Cerebral‑Voice Squad is more than a project management tool—it’s a voice‑driven collaborative coding ecosystem. By combining natural conversation with specialized AI agents, it empowers teams to move faster, reduce friction, and focus on creativity. Software projects evolve in real time, guided by dialogue, and supported by a secure, scalable cloud backbone.

ArchiWalk is a web application that transforms architectural floor plans into immersive first-person video walkthroughs. It works by: Uploading a Plan: Users upload an image of a floor plan. Using Veo: It utilizes Google's Veo video generation model (veo-3.1-fast-generate-preview) to interpret the 2D plan. Generating a Walkthrough: Based on the user's prompt (e.g., "start at the stairs"), it generates a realistic, simulated 3D video tour of the space. Customization: Users can quickly visualize any blueprint with full color, detail and with realistic atmosphere.

Drug valuation is paralyzed by disconnected teams using fragile, static spreadsheets that cannot keep pace with the complexity of drug development. This fragmentation traps Clinical, Commercial, and Strategy insights in opaque formulas and version control errors, leading to high-stakes decisions based on stale data. Synapse solves this by unifying stakeholders in a single workspace where AI agents instantly translate natural language conversations across functions into live, executable code and market models, creating a transparent, real-time single source of truth for key assets.

A better Excel :). It allows us not only to analyze spreadsheets but also use formulas, reference sheets in formulas. Most importantly we can use the agent to verify assumptions, check online for latest news to update the projections, etc

🍌 Nono Banana: A Controllable Benchmark for Non-English LLM Text Recognition LLMs still struggle to reliably extract non-english text from images. This has become a huge blocker for existing models, as over a billion people read and write Mandarin alone, yet most evaluation data for this task is tiny, messy, and english-centric. The core issue is simple: there’s no controlled, scalable way to test how LLMs behave on complex scripts. The existing datasets are scraped, mislabeled, inconsistent, and impossible to tune. you can’t say ‘make this 20% harder’ or systematically test radicals, stroke density, angles, blur, or font variation. Nono Banana fixes that. The name is a small RL wink — you keep saying ‘no no’ until the model improves — but the tech is the serious part.

AI Agent that reproduces the bug in the browser and generates a report when a Linear ticket it assigned or a verification is performed whenever a PR is raised on Github.

Probelm Large codebases are hard for both developers and AI tools to understand. Keyword search misses meaning, relationships between files are scattered, and navigating imports, calls, and dependencies is slow and error-prone. LLMs lack the structural and semantic context needed to give accurate suggestions. Full rescans don’t scale as repositories grow. Probabilistic heuristics rather than deterministic parsing can lead to subtle bugs and missed references Solution Hugging Tree builds a semantic knowledge graph of your codebase by combining graph traversal like follow imports, calls, and dependencies with vector search. It scans incrementally using git, and provides rich “context trees” that give LLMs deep awareness of structure, semantics, and relationships—resulting in precise code understanding, navigation, analysis, and planning.

Pair.AI is a real-time, conversational coding companion that feels like hopping on a video call with a senior engineer — except it’s always available, understands your entire screen, and never gets tired. You enter a clean, Google-Meets-style room with just two participants: you and your AI pair programmer. Share your screen. Speak naturally. Ask questions about your code. Request changes. And Pair.AI handles the rest. Using live speech-to-text and a multimodal AI model, Pair.AI listens to your voice, analyzes your code visually, and understands the exact context of what you're working on. Instead of pasting snippets or explaining everything manually, you can just talk out loud as you build.

AutoDrip is a React-based web application that utilizes Google's Nano Banana to enhance user-uploaded images for maximum social media appeal.

Cyber security attack simulation using RL and MCP

An app that allows you to code by voice entirely hosted in Visual Studio Code.

anoArchitect” is your complete design-to-render platform, built to turn ideas into architecture with precision and speed — powered by Nano Banana Pro. Core Concept: • Start with your vision: either a quick description of a space (e.g., “L-shaped living room with large windows and a mezzanine”) or an upload of a sketch or blueprint image. • Instantly generate a clean, editable SVG blueprint using either: • A Fast Assembly mode (leveraging a curated SVG component catalog) • Or a Full Generative mode using Nano Banana Pro to create the blueprint from scratch. • Take that blueprint and feed it into Nano Banana Pro to render multiple high-fidelity perspective views — top-down, isometric, interior eye-level — with control over lighting, camera, style, resolution.

**The Problem:** Standard AI coding agents are dangerous for Infrastructure-as-Code. They suggest deleting "idle" resources without understanding hidden dependencies, leading to production outages. > > **The Solution:** UnicoFinOps is an **MCP Server** that builds a real-time **Dependency Graph** of your infrastructure. It blocks dangerous operations *before* they happen, then uses **Google Gemini 3** to architect rightsizing strategies that save ~90% of spend safely.

Soulcaster turns noisy production complaints and user requests into one-click code fixes. It listens to real-world signals from Reddit and Github, automatically groups similar reports into clear, human-readable issues, and gives developers a single “Generate Fix” button that opens a ready-to-review GitHub pull request. Instead of hunting through forums, triaging duplicates, and wiring up boilerplate patches, teams get a live backlog that writes its own fixes and feature requests for them so that engineers can spend their time deciding what to ship, not deciphering what matters. For both open source and their own projects. https://soulcaster-gamma.vercel.app/

Axion is a constitution driven multi agent engineering system that lets AI work inside real codebases with structure and safety. Modern AI tools generate code but do not understand architecture or hidden rules. Axion fixes this with a Constitution for Codebases, a formal rule set that guides every agent decision. Axion uses a team of expert agents that collaborate like real engineers. They review each other, flag risks, call in specialists, and let the user join the conversation at any time. A sandbox environment enforces rules, tracks state, and ensures only safe diffs reach the repo. Axion solves Problem Statement 1 by giving models persistent context through the constitution and by assigning expert roles so agents reason with stable knowledge of the codebase. It solves Problem Statement 3 by enabling real multi agent and human collaboration. Axion turns natural language tasks into clean, architecture aware code changes.

This is a multiplayer code collaboration platform where multiple developers can work together in real-time on Linear tickets, with an AI agent that generates implementation plans and code changes, then automatically creates GitHub pull requests. The system features a chat interface for collaborating with teammates and AI, displays live diffs of proposed changes, and includes a PR management view—all synced through Supabase with polling-based real-time updates.

Aether Studio is where VS Code meets Google Docs, but with AI as an active teammate, not just a chatbot. It is a fully collaborative, agentic environment inside your browser.

Heimdall is an autonomous SRE engine that turns your system architecture into a live, intelligent dependency graph. When an outage occurs, localized diagnostic agents traverse the graph, stream their reasoning, and surface the critical failure path. A master Insight Engine then generates a precise, LLM-powered root-cause analysis with clear, actionable remediation.

Automated Reinforcement learning sandbox for security testing , and refining models approach to solve complex security vulnerabilities.