Inspiration
Modern software and data pipelines often suffer from the latency and excessive cost of traditional LLM orchestration. We built SqAr to solve this bottleneck. We wanted a flexible, robust agentic architecture that doesn't just generate text, but reliably breaks down complex developer workflows, writes and validates logic, and executes multi-step tasks autonomously all while running on high-speed, open infrastructure.
What it does
SqAr acts as an intelligent workflow copilot and autonomous agent: Task Decomposition & Reasoning: Evaluates user requirements, splits complex technical goals into verifiable subtasks, and dynamically plans execution. Tiered Model Routing: Routes deep logical queries and critical architectural planning to heavy reasoning models, while offloading routine validation, syntax formatting, and fast tool executions to lighter, ultra-low-latency models. Self-Correction & Execution: Automatically catches execution flaws, iterates over generated code/data artifacts, and delivers reliable, end-to-end results.
How we built it
Nebius Token Factory & AI Cloud: Deployed as our primary inference and execution backbone. We took advantage of low-latency endpoints to keep agent loops responsive and token-efficient. NVIDIA Models: Powered by NVIDIA Nemotron open-source models: NVIDIA Nemotron Ultra: Employed for deep reasoning, prompt decomposition, and zero-shot planning. NVIDIA Nemotron Nano: Utilized for lightweight classification, intermediate parsing, and fast conversational loops. Architecture: Developed with Python backend, utilizing serverless worker concepts for asynchronous agent task execution and real-time state management.
Challenges we ran into
Balancing response latency with reasoning accuracy in multi-turn agent loops.
Accomplishments that we're proud of
Successfully running end-to-end task execution autonomously with built-in self-correction.
What we learned
How effectively tiered orchestration between lightweight and deep-reasoning Nemotron models optimizes GPU compute credits without degrading output quality.
What's next for SqAr
Expanding native IDE extensions and terminal CLI tools for zero-friction developer adoption.
Log in or sign up for Devpost to join the conversation.