Roadmap

Where things stand

Two tiers are complete.

Production hardening. Cross-run dedup in the candidate queries, advisory locks around billed runs, a Postgres-marked test suite, a run and cost ledger with a heartbeat, per-task model routing, a Slack listener allowlist, and the agent verify-before-draft gate (ADR 0016).

Conversion intelligence. Funnel instrumentation, the worklist, the warm-path finder, freshness urgency, keyword-gap scoring on drafts, the interview prep pack and question bank, the rejection autopsy, and follow-up sequencing — all now measured through the conversion ledger (ADR 0022).

What is next

Persistent per-source health metrics. Pass statistics are ephemeral today: a source that has been quietly returning nothing for three weeks looks the same as a quiet week. A source_runs table and a history command would make the degradation visible. Half a day.

A comp overlay for undisclosed bands. Most postings do not disclose, so the comp gate has nothing to act on and the fit score loses a real signal. Backfilling from a public comp dataset would feed both. A day.

More agent tools. Reputation, recent news, engineering-blog recency — each is the kind of thing that decides whether a company is worth an application and that the pre-filter cannot see. Half a day each.

Cross-source dedup persistence. The canonical hash suppresses duplicates within a run but not across runs, so the same role reappearing next week is a new row.

Auto-promotion of discovered companies. karani discover finds ATS boards and promotes hits; the next step is promoting them into the curated list from measured yield rather than by hand.

A structured tool-call log per row, so an agent-mode verdict can be audited rather than summarised.

What is explicitly not on it

• Fine-tuning. Memory is few-shot context plus measurement, and the verdict loop is the eval. A fine-tune would be a worse version of both with a training pipeline attached.
• Submitting applications. Not a roadmap item, not a future consideration (ADR 0012).
• Scraping LinkedIn or Indeed. Hostile to their terms, and brittle.
• A dashboard. The terminal, one HTML file, markdown and Slack.
• Multi-tenancy. karani is one person’s tool and the schema says so.

Good first issues

The hardening tier is complete, so the remaining self-contained work is in the signal-quality tier: the per-source health metrics table and the additional agent tools. Both are well-bounded, both have acceptance criteria written down, and neither requires understanding the billed paths.

Read CLAUDE.md  first — the non-negotiable rules in section 4 apply to any change, and several of them are the kind of thing that looks like a style preference until you know which bug it came from. Conventions is the rest.