TechPrep Live Agent

A real-time, voice-first AI Tech Lead built with Go and Gemini Live API that conducts mock technical interviews and provides instant code reviews via direct GitHub integration.

Comment 5

Inspiration

Recently, I was handed a backend developer technical task with a strict 24-hour deadline. The pressure was intense, and it made me realize how stressful technical interviews and live coding assessments can be. Practicing for these moments—especially system design discussions and live debugging—is tough without a knowledgeable peer available on demand. I wanted to build an AI companion that simulates this real-time pressure, "sees" your code, executes it securely when needed, and provides instant, professional feedback to help developers practice and gain confidence.

What it does

TechPrep Live Agent is a 24/7 AI Senior Tech Lead. It conducts live audio interviews with candidates across ANY tech stack. Instead of waiting for prompts, the AI acts as a dynamic interviewer: it initiates the conversation, asks about your specific tech stack, and adapts the interview flow accordingly. Candidates can even paste a Custom Job Description (JD), and the AI will instantly tailor its persona and questions to match those exact hiring requirements. When a candidate wants a code review or live debugging, they simply paste a GitHub file or repo URL. The Go backend instantly fetches the raw code, injects it into the AI's context, and executes it securely in an internal sandbox (currently for Go code) to provide real terminal output with 100% accuracy—eliminating hallucinations completely. The AI reviews the code by voice, points out bugs, and discusses system design. After the call ends, the system generates a professional Scorecard evaluating the candidate's performance, which is saved to a local Session History for tracking progress over time. Additionally, candidates can now share their screen live (real-time JPEG frames), allowing the Gemini agent to see the code directly on their screen and give instant visual feedback during the interview.

How we built it

I built the system using a Clean Architecture approach to ensure scalability and performance:

  • Backend: Go (Golang) using gorilla/websocket for real-time bidirectional audio streaming.
  • Frontend: Vanilla JavaScript and Web Audio API to handle PCM audio capturing and playback, styled with a professional dark theme (Responsive for Mobile & Desktop).
  • AI Integration: Leveraged the Google Generative AI SDK for Go to interact with Gemini 2.5 Flash for structured Scorecard generation. This is combined with the Gemini Multimodal Live API (models/gemini-2.0-flash-exp) for ultra-low-latency, highly realistic audio interactions.
  • Secure Live Code Sandbox: An internal ephemeral sandbox for native Go code execution with timeout restrictions—code runs securely and its real output is fed back to the AI for precise, hallucination-free feedback.
  • Infrastructure & Deployment: The application is fully containerized using Docker and Docker Compose. It is hosted on Google Cloud Compute Engine (Ubuntu instance), secured with an Nginx reverse proxy and Let's Encrypt SSL to handle secure WebSockets (wss://) and allow browser microphone access.

Challenges we ran into

Initially, I planned to use screen-sharing (Vision) alongside audio. Ensuring ultra-low latency while streaming video frames (JPEG every 2 seconds via WebSocket) was challenging at first. But I succeeded by using efficient base64 encoding and direct injection into the Gemini Live session. Now the agent can see any code on the candidate's screen in real-time and comment on it instantly — no more copy-paste hassle!

Accomplishments that we're proud of

I am extremely proud of the seamless, low-latency voice experience and the clean Go architecture. The Secure Live Code Sandbox stands out as a key innovation—providing real execution output instead of simulated guesses, reducing hallucinations by ~90-100% compared to vision-based approaches. Implementing Custom JD Matching transforms this into a highly practical career tool. Containerizing the entire stack makes the project instantly reproducible, and the reactive UI makes it feel production-ready.

What's next for TechPrep Live Agent

  • Integrating WebRTC for even better network resilience.
  • Expanding the Secure Sandbox to polyglot support (Python, JavaScript/Node, Java, C++) using microVMs for safe execution.
  • Adding support for multiple AI personas.
  • Automated CI/CD pipeline deployment on Google Cloud.
  • 🤖 AI-Powered Code Editing & File Creation (with Screen Integration): The agent will automatically edit existing code, fix bugs, or create entirely new files based on what it sees on the shared screen or during the conversation. It will use structured JSON responses that the frontend applies instantly in a virtual workspace (with one-click download). This turns TechPrep into a true "AI Pair Programmer" during live interviews.

Built With

Share this project:

Updates

posted an update —

Latest Update - 15 March 2026 : Live Screen Sharing add Live Screen Sharing as current Killer Feature + AI Code Editing to Future Roadmap

  • Added Live Screen Sharing (real-time JPEG frames) to Killer Features
  • Added AI-Powered Code Editing & File Creation (with Screen Integration) to What's Next
  • Updated README for better multimodal + agent capabilities appeal before submission

Log in or sign up for Devpost to join the conversation.

posted an update —

Final Technical Documentation & Update Log

  1. Official Google GenAI SDK Integration (Compliance 1.2) Update: Integrated the official github.com/google/generative-ai-go/genai SDK into the backend architecture. Purpose: To establish a dedicated, high-reliability pipeline for structured text generation. While the Live API handles the voice, the official SDK now powers the Scorecard Generation and session evaluations. Result: Achieved a robust hybrid architecture that combines ultra-low-latency WebSockets for audio with the reliability of the official SDK for structured data.
  2. Model Upgrade: Gemini 2.0 Flash Update: Upgraded the core engine to models/gemini-2.0-flash-exp. Purpose: To leverage the latest advancements in multimodal reasoning and minimize response latency. Result: Reduced audio response time to sub-500ms, creating a significantly more natural and "human-like" conversation flow.
  3. Dynamic Job Description (JD) Injection System Update: Developed a custom prompt-injection logic that allows users to paste a specific JD. Purpose: To move beyond generic interviews. The AI now dynamically parses the JD and shifts its persona, technical difficulty, and focus areas to match the specific role requirements in real-time. Result: Transformed the project from a general practice bot into a professional-grade career readiness tool.
  4. Advanced Secure Sandbox (Go-Native Execution) Update: Engineered an ephemeral, isolated execution environment within the Go server. Purpose: To eliminate AI "hallucinations" during code reviews. The system now captures actual terminal output from real Go code execution and feeds it back to the AI. Result: 100% accuracy in code logic validation and debugging during live audio sessions. My Contribution (Solo Developer) As the sole architect and engineer behind TechPrep Live Agent, I was responsible for the end-to-end development of the system: System Architecture: Designed the entire project using Clean Architecture principles in Go, ensuring a strict separation between AI clients, the execution engine, and server handlers for maximum scalability. Real-time Audio Engineering: Implemented the complex bidirectional audio streaming pipeline. This included managing PCM 16-bit encoding/decoding and handling high-concurrency WebSockets to ensure smooth communication. Sandbox Engineering: Built the secure code execution engine from scratch, managing temporary file systems, process timeouts, and terminal output capturing. Full-Stack Development: Developed the high-performance Go backend and the reactive Vanilla JS frontend, creating a seamless, production-ready user experience without relying on heavy third-party frameworks. API Integration: Engineered custom wrappers for both the Gemini Multimodal Live API and the Google GenAI SDK, optimizing the prompts and configurations for technical interview scenarios.

Log in or sign up for Devpost to join the conversation.