Best-of-N selection
hone-select
Pick the best of several model outputs without letting one judge decide alone.
Open source · honeworks
Small tools for reliable AI workflows.
Read the code. Break the assumptions.
Five Python packages from my own AI pipelines. I wrote the specifications and acceptance tests; an AI coding agent wrote the code under my review, and a review bot checks every pull request. Each one works alone and fixes one failure I kept hitting.
- 5 public repositories
- Apache-2.0
- Alpha · 0.1.0
- Install from GitHub
Taste scoring
hone-taste
Estimate what people would like, and say whose taste the estimate came from.
Coming next
Tools people keep asking for, where today's options are scripts or one-off hacks.
-
Will be published soon
n8n-replay
Replay failed n8n executions in bulk: filter by workflow and time, group failures by error, dry-run first, replay at a safe rate through the public API, and keep an audit log of what was replayed and what happened.
Why this one (2 public requests)
People have asked for bulk retry in the n8n community since 2020; today it means one-by-one clicks or scripts against an internal endpoint.
-
Will be published soon
vram-steward
One coordinator for local AI servers sharing a GPU: VRAM leases across Ollama, ComfyUI and your own scripts, unloads that don't hang, and a priority queue so interactive requests don't wait behind batch jobs.
Why this one (3 public requests)
Mixing Ollama with image or audio models on one card leads to hangs and out-of-memory crashes; the fixes today are single-purpose unload nodes and scripts.
-
Will be published soon
batch-ledger
A reconciliation ledger for OpenAI and Anthropic batch jobs: match every result by custom_id, flag truncated and errored items, resubmit only what failed or expired, and report cost per job.
Why this one (2 public requests)
Batch APIs report a whole job as ended while individual items fail, expire or come back truncated; teams keep building their own reconcilers.
How they fit together
None of them imports another. They share one record format, so a hone-flow step can call hone-models, pick a winner with hone-select, and leave records hone-lens can analyze.
How they were built
An AI coding agent built each package from my written specification and acceptance tests. Design changes wait for my approval, and a review bot checks every pull request.
Need these controls in a real system?
Start with the failure your team cannot leave invisible.
The packages show how I build. The work is fitting the same ideas to your process, data and ownership.