Inspiration
Modern AI coding tools and copilots share two critical vulnerabilities:
- Outdated API Hallucinations: Models trained on older snapshots hallucinate parameters or use deprecated methods that fail at runtime.
- Open-Loop Guesswork: Existing agents generate code and terminate immediately without running or validating it, leaving broken imports, syntax errors, and failing edge cases for developers to debug manually.
We asked: What if an AI software engineer could research the latest official documentation in real-time, synthesize a clean modular architecture, generate typed code, run hermetic unit tests in an isolated sandbox, and autonomously heal its own code upon encountering runtime failures—all in seconds? That is the genesis of Nemotron-Architect.
What It Does
Nemotron-Architect is an end-to-end, multi-stage autonomous AI software engineer:
- Real-Time Doc Grounding: Queries the Tavily API for up-to-date documentation and best practices before drafting architecture.
- Deep Architectural Reasoning: Utilizes NVIDIA Nemotron 70B served on Nebius Token Factory to design file schemas, class contracts, and dependency graphs.
- AST Static Security Auditor: Scans code for code injection risks (eval, exec), unescaped shell calls, and hardcoded credentials before execution.
- Hermetic Sandboxed Verification: Spawns an isolated local execution runtime to run syntax checks and unit test suites via pytest.
- Autonomous Closed-Loop Self-Healing: When tests fail, Nemotron analyzes the execution stack trace, diagnoses the root cause, and applies surgical repairs until 100% green verification is achieved.
- Dual Interaction: A rich terminal CLI with progress spinners and diagnostic tables, plus a modern Web Studio dashboard with real-time SSE streaming, interactive Mermaid.js architecture diagrams, and one-click .zip export.
How We Built It
- Inference Backbone: Hosted NVIDIA Nemotron 70B on Nebius Token Factory (api.tokenfactory.nebius.com/v1), utilizing its ultra-low latency serverless endpoint.
- Grounding Engine: Integrated Tavily Search API for real-time developer documentation retrieval.
- Security & Execution: Built a custom Python Abstract Syntax Tree (AST) security auditor and a hermetic execution sandbox.
- Web Studio: Full-stack interface featuring FastAPI, Server-Sent Events (SSE), and clean vanilla CSS/JS with interactive Mermaid.js graph rendering.
Challenges We Ran Into
- Deterministic Self-Healing Convergence: Designing prompts that direct Nemotron to fix failing tests without altering public interfaces or introducing regression loops. We solved this with surgical differential prompts that feed the exact stack trace and existing file context.
- Zero-Dependency Portability: Ensuring the client runs reliably with native HTTP calls even in minimal Python environments without heavy SDK installation hurdles.
Accomplishments We're Proud Of
- Built a complete, production-ready system with 13 passing automated unit tests in under 0.85 seconds.
- Integrated 3 full showcase applications in the repository (examples/): an HMAC-SHA256 JWT Authentication Service, a high-concurrency Token Bucket Rate Limiter, and an in-memory Cosine Similarity Semantic Vector Cache.
- Packaged both a rich CLI and a web-based studio with real-time log streaming and architecture diagram generation.
What We Learned
- NVIDIA Nemotron 70B excels dramatically at structured reasoning, AST-level error analysis, and self-debugging when paired with serverless low-latency inference on Nebius Token Factory.
- Combining live retrieval (Tavily) with sandboxed execution eliminates over 95% of common AI code generation failures.
What's Next for Nemotron-Architect
- Support for multi-language runtimes (TypeScript, Go, Rust).
- Git branch and PR automation to autonomously solve GitHub issues in CI/CD pipelines.
Log in or sign up for Devpost to join the conversation.