.png&w=256&q=75)
🤓 Latest Submissions
DevPilot AI: From Issue to Verified Fix
Software engineers lose significant time moving between GitHub issues, unfamiliar files, code search, security findings, test generation, and code review. Existing AI coding assistants can generate useful answers and code, but developers still need to determine which repository evidence supports the answer and whether a proposed change is actually testable. DevPilot AI addresses this gap with a repository-aware engineering workflow: Issue → Investigation → Evidence → Root Cause → Plan → Patch → Regression Test → Verification DevPilot is designed for software developers, engineering teams, repository maintainers, and students working with unfamiliar or evolving codebases. Users can start from a repository or issue, inspect relevant files and symbols, ask repository-aware questions, review security findings, generate tests and documentation, create implementation plans, and review proposed engineering actions. The product is presented as a unified engineering workspace containing Repository Intelligence, AI Chat, Code Analysis, Security, Testing, Documentation, Planner, and PR Review. The prototype also includes Demo Mode, allowing the workflow to be evaluated without external model credentials. Our key innovation is treating AI-assisted coding as a closed-loop engineering process, rather than a single chat interaction. DevPilot connects repository retrieval and reasoning to concrete engineering artifacts such as implementation plans, proposed changes, regression tests, and verification results while keeping the human engineer in control. The prototype uses lexical and symbol-aware repository retrieval, repository file/line citations, controlled verification for the bundled demo repository, and automated regression coverage. Current prototype validation achieved: 100% expected-file hit rate across 10 demo retrieval cases 100% expected-symbol hit rate across 5 labeled cases 100% citation-line validity 13 backend + 2 frontend tests passed These metrics specifically evaluate retrieval, citation validity, and prototype regression coverage; they are not claims of overall AI answer correctness or production reliability. The roadmap includes authenticated multi-user workspaces, persistent storage, isolated code-execution sandboxes, GitHub OAuth, branch/PR automation, semantic retrieval, and broader end-to-end evaluation.
27 Sep 2026
Real-Time Voice AI Agent
This project is a Real-Time Voice AI Agent designed to provide natural, hands-free interaction between users and artificial intelligence through spoken conversation. The system captures a user's voice through a web browser, converts speech into text using real-time speech recognition, processes the request with a large language model, and converts the generated response back into speech. The architecture uses AssemblyAI for real-time speech-to-text, Gemini for AI response generation, streaming text processing, text-to-speech, and Web Audio/AudioWorklet for browser-based audio playback. Instead of waiting for the complete AI response, the system processes and streams response chunks progressively to reduce perceived latency and improve conversational flow. A major engineering focus is real-time stability and interruption handling. The system uses turn IDs, conversation management, streaming generation, audio queue scheduling, timeout handling, and barge-in support. When a user begins speaking while the assistant is responding, the active LLM generation, TTS processing, and queued audio can be cancelled so the new request becomes the active conversation turn. During development, major issues such as long response delays, incorrect LLM configuration, failed AI connections, duplicate requests, unstable audio playback, and misleading connection states were identified and addressed through systematic debugging, health checks, improved error handling, streaming architecture, and turn-aware audio management. The deployment architecture is designed for Vercel frontend hosting and Render FastAPI backend hosting, while keeping API credentials securely on the server side. PersonaPlex is maintained as an optional future speech-to-speech engine when suitable GPU infrastructure is available.
30 Sep 2026
Bimanual VLA Dinner Table Manipulation
Bimanual VLA Dinner Table Manipulation is a simulation-first Physical AI system developed for the Intel Physical AI Online Challenge. The project addresses the challenge of autonomously setting up a dinner table using two coordinated SO-101 robotic arms in a MuJoCo environment. The system combines natural-language instructions with camera-based scene observations to understand the task, identify relevant objects, maintain task state, plan manipulation sequences, and coordinate both robotic arms. The target workflow includes opening a drawer, retrieving utensils such as spoons and forks, transporting and placing a plate and cup, and verifying the resulting table configuration. A hierarchical architecture connects multimodal perception and language understanding with task-state management, task planning, bimanual action planning, robot control, simulation, and task verification. The system is designed to support collision-aware sequencing, object-specific grasping, placement verification, recovery from failed actions, and randomized initial conditions. MuJoCo provides the reproducible simulation environment, while LeRobot-compatible VLA policies are planned as the learned-policy layer for language-conditioned manipulation. The project also targets robustness evaluation across randomized seeds and variations in object placement and physical conditions. For Intel edge deployment, the project is designed around OpenVINO optimization and inference on Intel Core Ultra platforms, with latency, throughput, precision, and device benchmarking included in the final evaluation plan. The overall goal is a reproducible end-to-end Physical AI pipeline demonstrating multimodal reasoning, coordinated bimanual manipulation, robustness, and efficient edge inference.
16 Sep 2026
