Caret Meetup #2: Talks and Videos

Caret held its second meetup and shared five talks. We aim for AI as a coding partner, not just a tool. Below are the overview video and session details.

Overview video:

Open-source Caret & Semantic Merging

• Speaker: Byungseok Yang (CTO)
• Story: Forking cline (VSCode extension) while surviving merge hell.
• Lessons:
  • Started with a 100% compatible fork.
  • First big merge failed—.cline backups confused AI between Caret and cline code.
  • Adopted Minimal Invasion: avoid touching upstream files; use wrappers to cut merge cost.
  • Authored detailed Caret Rules so AI understands architecture/locations/conventions.
  • Realistic AI collaboration: AI forgets context; developers must monitor and intervene.
  • Vision: humans set strategy and verify; AI handles heavy implementation while keeping up with upstream. Session video:

Persona Image Generation with Nanobanana

• Speaker: Ki-Hwan Kim (CEO)
• Goal: Let users create custom AI avatars.
• Path: Built from PRD with AI tools; initial attempts failed (only sample images).
  Found a working OSS project (nanobanana-ads), showed the code to the agent, and succeeded immediately.
• Lesson: Concrete, working examples boost AI implementation accuracy far more than abstract requirements. Session video:

Integrating Caret with LiteLLM

• Speaker: Donghak Kim (Contributor)
• Problem: Many LLMs with different APIs; hard to switch models and manage costs.
• Solution: Chose OSS LiteLLM (“LLM ORM”) to normalize 100+ models to OpenAI format.
• Features: auto-failover (GPT-4 → Claude), load balancing, real-time cost tracking.
• Frank take: UI/UX is rough but speed to customers mattered.
• Caret’s build: “Caret Router” + social login + gRPC backbone. New users get $10 credits to try models. Session video:

Semantic Coding Model for TypeScript

• Speaker: Seungwan Oh (PhD candidate, Chonnam Univ.)
• Problem: Code LLMs trained on static text don’t truly grasp execution flow/semantics.
• Approach (“SemCode”): include execution traces, constraints, debugging cases alongside code.
• Shift: OSS-Instruct TS data quality was low; switched to vetted sources (e.g., LeetCode) to raise valid data yield. Session video:

AI Coding Assistant Codebase Strategy

• Speaker: Jaehoon Choi (Samsung)
• Framing: Two metaphors—Developer (analyze AST live; accurate, local, private) vs Librarian (pre-index with embeddings/RAG; fast concept search).
• Goal: Caret’s hybrid—add Cursor-like flexible search to cline-like precision. Handle both tight refactors and broad exploratory questions. Session video:

Caret is an open-source Korean Vibe Coding tool, running regular meetups with contributors. We welcome anyone interested in AI-based development to join and build together.***