Case study · located work
1k Papers
A live site that classifies 1,018 AI papers against 24 topics, with Jev doing the classification after an LLM has summarised each paper.
ObservedHassan El Mghari
- Problem
- A live site that classifies 1,018 AI papers against 24 topics, with Jev doing the classification after an LLM has summarised each paper.
- Jev’s role
- One typed classification per paper against a fixed 24-topic taxonomy; the LLM summary is Jev's input, not the classifier.
- Architecture
- Paper → LLM summary → Jev classification against 24 topics → browsable topic index.
- Why it matters
- A bulk-classification shape rather than an agent loop: the corpus is fixed, the taxonomy is fixed, and the interesting cost question is per-item. The build's cost and latency numbers are self-reported by the author and are not repeated here; only the published corpus size and topic count are.
Code boundary
Jev answers the bounded questions described above. Thresholds, retries, side effects, and anything that must be reproducible stay in the surrounding application code. The research record does not capture this project’s exact question set or threshold values; read the linked source before copying the architecture.
Sources
Limitations
- A located repository or demonstration does not establish production reliability or independent validation.
Reusable pattern and related records
- Decision sidecar PlausibleA typed judgment service sits beside ordinary code; code owns effects and safety boundaries.
- Local Activity Auto-Timeline Authored HypothesisA private desktop timeline deciding whether each app/window event starts, continues, or ends an activity.
- TypeSafe reports a low input-token price and no metered output-token charge for Jev. Vendor ClaimPublished pricing is echoed across launch discussion, but remains mutable vendor pricing.
- Parallel decision matrix PlausibleOne state is evaluated against many named questions, replacing repeated prompt/parse cycles.