Takoping

Takoping is a developer onboarding platform that provides repository visualization using genAI that will accelerate onboarding time for devs that struggle to understand complex codebases.

Comment

Inspiration

  • New developers lose days piecing together unfamiliar repo structures and hidden conventions.
  • Senior engineers spend time re-explaining patterns, causing inconsistent onboarding.
  • We wanted an AI teammate that “shows, not tells” the architecture, flows, and standards of any repo.

What it does

  • Interactive, zoomable repository visualizer with level-of-detail to reveal just enough at each zoom level.
  • Contextual code viewer and an intelligent chat that can highlight, open files/ranges, and “take you there.”
  • Agent-created diagrams/tabs on demand (e.g., feature data flow, architecture slice) tied to the answer.
  • Hybrid search: semantic code retrieval enriched with repository tree context.

How we built it

  • Client: Vite + TypeScript, React Flow for the canvas, Jotai for state, Shadcn UI for components.
  • Server: Bun runtime, Vercel AI SDK for tool-use/agent orchestration and provider routing.
  • AI: NVIDIA NIM LLM (llama-3 1-nemotron-nano-8B-v1) on EKS and Retrieval Embedding NIM on public nvidia endpoint.
  • Data: Hybrid Postgres (structure, presets) + ChromaDB (vectorized code chunks) with cosine similarity.
  • Integration: GitHub API ingestion; real-time UX via Bun WebSockets.

Challenges we ran into

  • Balancing context windows: stitching RAG results with tree-structured repo context without bloat.
  • Fast LoD rendering for large repos: avoiding DOM overload while keeping interactions snappy.
  • Making agent tools work with smaller parameter model
  • GitHub API constraints: rate limits, large trees, and respect for ignore rules during ingestion.
  • Clean schema boundaries between Postgres (hierarchy/presets) and ChromaDB (semantic content).
  • Deploying on eks with amazon

Accomplishments that we're proud of

  • Hybrid RAG + tree context that grounds answers in concrete code and structure.
  • Smooth zoom-aware LoD that keeps huge repositories usable and visually coherent.
  • Simple env-based switching between local dev and AWS NIM endpoints.
  • Deploying onto aws eks with full cloud stack

What we learned

  • Visual grounding dramatically improves trust and comprehension vs text-only answers.
  • LoD thresholds and virtualization are essential for performance at scale.
  • Agent tools should mirror real user actions to feel intuitive (navigation > narration).
  • Clear separation of structure (Postgres) and semantics (Chroma) simplifies the system.

What's next for Takoping

  • Fix auto-diagramming: on-the-fly, answer-specific diagrams/tabs that persist for reuse.
  • GitHub OAuth and webhooks for live updates; branch switching in the visualizer.
  • Architectural pattern detection and anti-pattern alerts with targeted refactor suggestions.
  • Guided, shareable, replayable “tours” for common tasks and repo walkthroughs.
  • Infra-as-code (Terraform) for VPC, SageMaker endpoints, and secure service wiring.
  • “Open in Cursor” flows and issue/PR creation for detected problems or improvement suggestions.

Built With

Share this project:

Updates