ReasonAI

ReasonAI is an AI-native engineering workspace to discover ideas, design and review architectures, collaborate in real time, and learn DSA visually. Powered by NVIDIA Nemotron on Nebius.

Comment

ReasonAI — Build Better Judgment

AI-Native Engineering Workspace | Nebius × NVIDIA Global AI Hackathon | Track: Best Apps & Agents

AI can generate answers in seconds. ReasonAI helps engineers develop the judgment to know what to build, why it works, and what to change.

Inspiration

Engineering today has an information abundance problem, not an answer shortage.

A developer discovers a new technology in one tab, asks an AI to explain it in another, sketches an architecture on a separate canvas, debates trade-offs in chat, and practices algorithms on an isolated coding platform. While individual AI tools are powerful, the engineering workflow is deeply fragmented. Most AI assistants optimize for producing an immediate answer; very few help an engineer interrogate that answer, translate it into a system design, collaborate on trade-offs, and build durable technical understanding.

We built ReasonAI around a single foundational question:

What if AI were embedded inside the engineering workflow—operating against real design and code artifacts—rather than sitting beside it as a disconnected chatbot?

ReasonAI is an AI-native workspace built for software engineers, architects, and technical learners. Its goal is not to replace engineering judgment, but to sharpen it through a unified Discover → Reason → Build → Master loop.

What it does

ReasonAI unifies technical discovery, collaborative system architecture, and algorithmic mastery into a single multi-platform environment across web and mobile.

1. Discover: Intelligent Knowledge Feed & Grounded Research

ReasonAI continuously curates an engineering Knowledge Feed. An automated backend pipeline ingests current technical sources, removes duplicates, validates imagery (persisted via Cloudflare R2), and uses NVIDIA Nemotron via Nebius Token Factory to enrich stories with structured engineering context: what happened, why it matters architecturally, and what to explore next.

  • Interactive Investigation: Users can open any story and Ask ReasonAI to verify claims, unpack underlying technologies, or compare architectural approaches.
  • Live Grounding: When questions require up-to-date external evidence beyond model weights, the agent invokes Tavily search and extraction to return source-backed answers.

2. Reason & Build: Context-Aware System Design Canvas

System design requires reasoning about traffic patterns, failure modes, state consistency, and operational bottlenecks—not just drawing boxes.

  • Artifact-Aware AI: In ReasonAI's Konva-powered System Design workspace, the AI reads the actual nodes, edges, and topology on the active canvas. Engineers switch between specialized reasoning modes—Chat, Review, Fix, and Eagle View—to evaluate scalability, spot single points of failure, and stress-test architectures.
  • Typed Canvas Proposals with Human-in-the-Loop Safety: Rather than stopping at prose advice, ReasonAI generates typed structural proposals for adding, modifying, or connecting components. Every proposal is validated against a strict application schema and presented for human review. No AI output mutates the canvas until the engineer explicitly inspects and accepts it.
  • Real-Time Multi-User Collaboration: Architecture reviews are team activities. A dedicated Go WebSocket collaboration service manages ordered operations, acknowledgements, user presence, event replay, and reconnect synchronization—allowing multiple engineers to co-design live while keeping AI proposals strictly gated behind the review boundary.

3. Master: 375-Problem DSA Explorer, Execution Studio & Code Intelligence

Coding practice should build problem-solving intuition rather than encourage solution memorization. ReasonAI provides a structured 375-problem DSA catalog integrated with an interactive execution and analysis environment:

  • Run / Submit & Results Studio: Engineers write code and use the integrated Run and Submit workflows to evaluate solutions against structured test cases. The Results Studio surfaces case-by-case execution outcomes, edge-case behavior, and provisional grading feedback.
  • Code Intelligence: Beyond pass/fail checks, ReasonAI provides Code Intelligence—analyzing submitted implementations for time and space complexity, structural patterns, algorithmic trade-offs, and concrete refactoring opportunities.
  • LangGraph Tutor & Validated Visual Walkthroughs: A LangGraph-orchestrated agent maintains persistent learning state and bounded tool loops to deliver progressive hints rather than spoilers. When illustrating data structures or state transitions, it emits schema-validated visual walkthroughs rather than brittle, arbitrary markup.
  • Native Mobile Experience (Flutter): A native Flutter Android application (downloadable APK) extends the Knowledge Feed, Ask ReasonAI, and DSA learning and practice flows to mobile devices.

How we built it

Our core architectural principle was to strictly separate what the model reasons about from what the application is allowed to mutate.

Layer Implementation Engineering Purpose
Model Inference Nebius Token Factory + NVIDIA Nemotron Powers feed enrichment, grounded research synthesis, system design critique, canvas proposals, DSA tutoring, and Code Intelligence.
Model Routing Nemotron Lightning / Super / Ultra + Auto Mode Matches workload to latency/depth requirements. In System Design, Auto routing dynamically escalates complex multi-constraint turns from Super to Ultra.
Agent Orchestration LangGraph, streaming SSE, bounded tool loops, persistent memory Coordinates multi-step reasoning, state persistence, and tool execution without runaway loops or context drift.
Search & Grounding Tavily API Supplies live web evidence and source citations when queries require current facts outside static training data.
Web & Interactive Canvas Next.js, React, TypeScript, Tailwind CSS, Konva Renders the interactive architecture canvas, node/edge overlays, DSA workspace, Results Studio, and Code Intelligence panels.
Safe Action Boundary Typed Proposal Contracts, Schema Validators, Reducer Preflight Parses model outputs into deterministic node/edge operations, resolves references against current state, and enforces human approval before canvas mutation.
Real-Time Collaboration Go (Golang), WebSockets, op-ack, replay buffer Synchronizes live multi-user canvas edits, cursor presence, and reconnection state independently of the Python AI backend.
Backend & Persistence FastAPI (GCP Cloud Run), Supabase PostgreSQL, Cloudflare R2 Manages server-side API credentials, user/judge sessions, 375-problem DSA catalog state, execution workflows, and validated media assets.
Mobile Client Flutter (Dart) Delivers a native Android APK for on-the-go Knowledge Feed exploration, Ask ReasonAI investigation, and DSA study flows.

Key Architectural Principle: AI proposes; the engineer decides; the application enforces deterministic rules. All model credentials remain strictly server-side, and stale or malformed model outputs are caught by preflight validators before reaching the UI state.

Challenges we ran into

  • Making AI canvas mutations operationally safe: Asking an LLM to critique an architecture in text is easy; letting it modify an active graph without corrupting existing node/edge relationships is hard. We solved this by designing typed operation contracts, reference resolution against live editor state, reducer preflight checks, and an explicit human accept/discard staging layer.
  • Hardening streamed tool calls across model tiers: Streaming responses that interleave reasoning tokens, prose, and structured JSON tool-call fragments can break naive parsers. We built resilient server-side stream assembly and validation layers to handle chunked tool invocations reliably across Nemotron tiers.
  • Synchronizing real-time collaboration with staged AI proposals: When multiple users edit a canvas simultaneously via WebSockets, an AI proposal generated seconds earlier might reference deleted or modified nodes. Our Go WebSocket service handles ordered operation acknowledgements and replay, while client preflight checks ensure accepted AI proposals reconcile cleanly against the latest committed state.
  • Delivering trustworthy visual tutoring and code feedback: Free-form LLM-generated HTML/SVG frequently breaks layouts or hallucinates steps. For DSA walkthroughs, Results Studio diagnostics, and Code Intelligence, we constrained model outputs to strict, validated schemas that render deterministically in the UI.

Accomplishments that we're proud of

  • End-to-end working product across Web & Mobile: A live, multi-service production web application (reasonai.tech) with instant Hackathon Judge access, paired with a downloadable native Flutter Android APK and MIT-licensed source code.
  • Closing the loop in DSA mastery: Expanding the DSA workspace to a structured 375-problem explorer complete with Run/Submit workflows, Results Studio diagnostics, Code Intelligence complexity/refactoring analysis, and LangGraph progressive tutoring. (Note: To maintain strict technical honesty, our Run/Submit grading and Code Intelligence provide rapid developer verification and AI-assisted feedback rather than claiming to be an officially certified competitive-programming judge or unmeasured educational benchmark.)
  • Verifiable AI-to-Canvas action boundary: Moving beyond passive chat so the assistant generates inspectable, schema-validated architecture modifications grounded in real node and edge IDs.
  • Dedicated Go real-time infrastructure: Engineering a standalone WebSocket collaboration engine with presence, acknowledgements, and reconnect replay rather than relying on fragile polling.
  • Disciplined multi-tier Nemotron routing: Leveraging Nebius Token Factory's OpenAI-compatible interface to route workloads across Nemotron Lightning, Super, and Ultra—optimizing the trade-off between interactive speed and deep architectural reasoning.

What we learned

Building a high-utility AI engineering tool is not about wrapping a prompt around a text box. The real engineering challenge lies in defining the boundary between probabilistic model reasoning, deterministic software validation, and human judgment.

We learned that AI agents become significantly more trustworthy when their outputs are treated as untrusted, typed proposals until verified by application logic and approved by the user. We also saw firsthand how matching model tiers to cognitive tasks—fast models for enrichment and quick hints, deeper reasoning models for architectural bottlenecks and Code Intelligence—dramatically improves the user experience.

What's next

Our roadmap focuses on deepening verification, automation, and ecosystem extensibility:

  1. Hardened Multi-Language Sandbox Execution: Upgrading the DSA Run/Submit pipeline with isolated, resource-metered container sandboxes for strict runtime benchmarking across additional languages.
  2. GitHub-to-Canvas Architecture Ingestion: Allowing engineers to point ReasonAI at a live repository to automatically map service boundaries and generate an initial editable System Design canvas.
  3. Automated Layout & Simulation Engines: Adding deterministic graph auto-layout alongside quantitative traffic/latency simulation overlays in System Design.
  4. MCP (Model Context Protocol) Agent Integrations: Exposing ReasonAI's canvas and code workspaces to external developer tools via standardized MCP interfaces.

Work Completed During the Hackathon vs. Prior Foundation

To ensure complete transparency for judges:

  • Pre-Existing Foundation: The foundational project structure and initial workspace concepts existed prior to the hackathon.
  • Built During the Eligible Hackathon Period: We substantially architected and shipped the core intelligence, execution, and collaboration layers during the hackathon window, including:
    • Full Nebius Token Factory + NVIDIA Nemotron integration across Lightning, Super, and Ultra tiers with Auto escalation in System Design.
    • LangGraph agent orchestration, bounded tool loops, and persistent conversation/learning memory.
    • The 375-problem DSA Explorer, Run/Submit execution pipeline, Results Studio, and Code Intelligence analysis engine.
    • AI Canvas Proposals with strict schema validation, reducer preflight checks, and human-in-the-loop approval, plus Chat, Review, Fix, and Eagle View modes.
    • The standalone Go WebSocket real-time collaboration service (presence, acks, replay, reconnect handling).
    • The automated Knowledge Feed discovery/deduplication/enrichment pipeline, Ask ReasonAI, and Tavily live research grounding.
    • The native Flutter mobile application (with shareable Android APK) for Knowledge Feed and DSA workflows, alongside one-click Hackathon Judge access and automated test coverage.

(See HACKATHON_SUBMISSION.md and the commit history in the repository for a granular file-by-file record.)

Recommended 3-Minute Judging Path

  1. Instant Web Access: Open https://reasonai.tech and click Continue as Hackathon Judge.
  2. Discover: Browse the Knowledge Feed, open an enriched engineering story, and use Ask ReasonAI to run a grounded Tavily research query.
  3. Reason & Build: Open System Design, switch between Review / Fix / Eagle View, and ask the agent to propose scalability improvements. Inspect and accept a typed canvas proposal, then test Live Share across two browser tabs to see the Go WebSocket collaboration service in action.
  4. Master: Open the 375-problem DSA Explorer, write or test a solution using Run / Submit, inspect case diagnostics in Results Studio, review complexity insights in Code Intelligence, and request a progressive hint or validated visual walkthrough from the tutor.
  5. Try Mobile (Optional): Install the Flutter Android APK to test the mobile Knowledge Feed, Ask ReasonAI, and DSA flows.

Constructive Feedback on Nebius Token Factory & NVIDIA Nemotron

  • What Worked Exceptionally Well: Nebius Token Factory's OpenAI-compatible API allowed us to unify five distinct engineering workloads (batch story enrichment, grounded research, graph-aware architecture critique, DSA tutoring, and Code Intelligence) behind a single clean server-side provider abstraction. Having access to NVIDIA Nemotron Lightning, Super, and Ultra made dynamic latency-vs-depth routing practical in production.
  • Constructive Opportunities for Improvement:
    1. Model Lifecycle & Streamed Tool-Call Documentation: Clearer deprecation notices, canonical model-ID discovery endpoints, and provider-specific documentation for handling streamed tool-call JSON fragments alongside reasoning/content fields would significantly reduce integration friction for stateful LangGraph agents.
    2. Image Generation Predictability: Because image-generation endpoint availability fluctuated during development, we engineered our Knowledge Feed to validate and persist source imagery in Cloudflare R2 rather than relying on live image generation.
    3. Proposed Feature — First-Party Nebius Token Factory MCP Server: We strongly recommend that Nebius introduce an official, first-party Model Context Protocol (MCP) interface for Token Factory. An official MCP server exposing live model catalog inspection, tier latency/cost telemetry, quota monitoring, and standardized tool-call testing directly inside IDEs and agent frameworks would make operating multi-tier Nemotron agents dramatically faster for developers.

ReasonAI — Build Better Judgment.
Live Web App: https://reasonai.tech | Android APK: Google Drive Download | GitHub: saksham60/recallstack

Built With

Share this project:

Updates

Submission history