Feed resume review edits through Avocet corrections pipeline #112

New issue

Open

opened 2026-05-05 13:17:59 -07:00 by pyr0ball · 0 comments

pyr0ball commented

2026-05-05 13:17:59 -07:00

Owner

Context

As of v0.9.2, the Resume Review modal lets users edit the proposed summary and experience bullets before accepting them. Each edit is a (proposed, accepted) pair: the LLM output and what the user actually wanted.

Goal

Feed these correction pairs into the Avocet corrections pipeline so they can be used to fine-tune the resume optimizer LLM.

Data shape

When a user edits and approves:

Summary: { proposed: string, accepted: string } — a text_diff correction
Experience bullets: { proposed: string[], accepted: string[] } — a bullets_diff correction

Both can be extracted from the decisions dict that the /review endpoint already receives.

Implementation sketch

In dev-api.py POST /resume_optimizer/review, after applying decisions, check if any edited_text or edited_bullets differ from the original proposed values.
If so, write correction records to a resume_optimizer_corrections table (job_id, section, proposed_json, accepted_json, created_at).
Export endpoint or Avocet import hook to turn these into SFT (supervised fine-tuning) candidates.
Wire into Avocet label queue for human review before training inclusion.

Notes

The correction only fires when accepted=true AND the text was modified (proposed != accepted).
Rejections (accepted=false) are not corrections — they revert to original and carry no training signal.
This is the same Deterministic Bootstrap pattern used in Peregrine job_quality and Snipe listing_trust: high-confidence accepted edits self-certify; uncertain or sparse edits go to Avocet queue.

Labels

backlog
fine-tuning
avocet-integration

## Context As of v0.9.2, the Resume Review modal lets users edit the proposed summary and experience bullets before accepting them. Each edit is a (proposed, accepted) pair: the LLM output and what the user actually wanted. ## Goal Feed these correction pairs into the Avocet corrections pipeline so they can be used to fine-tune the resume optimizer LLM. ## Data shape When a user edits and approves: - **Summary**: `{ proposed: string, accepted: string }` — a text_diff correction - **Experience bullets**: `{ proposed: string[], accepted: string[] }` — a bullets_diff correction Both can be extracted from the `decisions` dict that the `/review` endpoint already receives. ## Implementation sketch 1. In `dev-api.py` POST `/resume_optimizer/review`, after applying decisions, check if any `edited_text` or `edited_bullets` differ from the original proposed values. 2. If so, write correction records to a `resume_optimizer_corrections` table (job_id, section, proposed_json, accepted_json, created_at). 3. Export endpoint or Avocet import hook to turn these into SFT (supervised fine-tuning) candidates. 4. Wire into Avocet label queue for human review before training inclusion. ## Notes - The correction only fires when `accepted=true` AND the text was modified (proposed != accepted). - Rejections (accepted=false) are not corrections — they revert to original and carry no training signal. - This is the same Deterministic Bootstrap pattern used in Peregrine job_quality and Snipe listing_trust: high-confidence accepted edits self-certify; uncertain or sparse edits go to Avocet queue. ## Labels - backlog - fine-tuning - avocet-integration