Jeongki Park

Native multimodal at 0.8B is genuinely impressive - most teams trade off size for capability, but 262K context windows + text/image/video in under 1B parameters changes the edge deployment math. The 9B beating GPT-OSS-20B on GPQA Diamond is interesting. Curious about structured output reliability at 0.8B though - small models tend to drop JSON schema adherence under complex instructions. Is...

This hits a real pain point. As a solo founder, I ship PRs every day but the changelog and social updates always fall behind. The gap between "work done" and "work communicated" keeps growing. The brand voice learning is what caught my attention - most AI content tools sound generic out of the box. How many sample pieces does it typically need before the output starts matching your actual tone?...

Worktree isolation per agent is the key design choice here. Without it, parallel agents would constantly step on each other's changes. The persistent daemon surviving crashes is a nice touch too. How are you handling merge conflicts when multiple agents edit overlapping files?

The eval-driven approach makes sense. Most teams copy skill files across projects and hope they still work after a model update - there's no feedback loop telling you the context degraded. Having structured evals that catch regression before it hits production is the missing piece. Curious about the version compatibility matrix. When a new model version drops (say Claude Opus to Sonnet), how...

Behavior-based scoring is the right call. Most registry security tools just check known CVE lists, but the real danger is packages that pass all the obvious checks and do something unexpected at install time. Focusing on what the code actually does rather than what the listing claims is a much stronger signal. The IDE extension scanning installed extensions in real time is a nice touch - most...

The cross-agent translation problem is real. I have 19 skill files for Claude Code and every time I try something in Cursor the format is completely different. Having a single source that compiles to each agent's format would save a lot of duplicated effort. How does Primer handle codebases with multiple languages? Does it generate separate skills per language or unified ones?

The plan-stage validation approach is really smart. Most governance tools catch problems after code is written, by then the developer already invested time and pushes back on changes. Catching it during the planning phase is a much better feedback loop. Curious about the ML-based rule matching - how does it handle edge cases where a task touches multiple domains with conflicting rules? Does it...