Roadmap
A dependency graph, not a calendar.
No dates. Every arrow below is a real dependency — feedback loops are structurally meaningless before execution is real, and specialization is premature before evaluation exists to measure whether "specialized" is actually better.
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
Training / evolution plan
Where "training" actually belongs — answered directly, not hedged
Retrieval + structured memory first
What's already built — cheapest, most interpretable, easiest to debug when it's wrong, and it's already producing real signal today via the Module 1 ↔ Module 3 loop.
Context / prompt specialization second
Org-specific terminology, severity conventions, and known-risky-area weighting encoded as retrieved context, not baked into weights. Still fully interpretable, still cheap to iterate.
Evaluation-driven iteration third
The actual gate. Without packages/evaluation being real, there's no honest way to know if steps 1–2 made anything better — only that they changed something.
Fine-tuning / preference learning last, and narrow
Only once there's real labeled trace volume — hundreds of real approve/reject/correct decisions per task type — and only for specific sub-tasks where retrieval and context genuinely plateau. Full-system fine-tuning is explicitly the wrong first move: expensive, hard to debug, and it throws away the interpretability retrieval-based memory gives for free.
What's missing, concretely
Feedback should be human-labeled not just as approve/reject (already captured) — but why, in structured form: wrong risk level? Missing test type? Tone mismatch with team convention? That's currently missing, and it's a real, concrete next feature — not vague future work.
Partnership vision
The foundation here is independently built and honestly scoped. The next real step — organization-specific specialization, real historical data, real feedback at volume — isn't something one person builds alone in good faith; it needs a real QA environment's data, workflows, and judgment.
If that's an interesting problem to you or your team, I'm open to technical collaboration, pilot work, or figuring out what a deeper partnership could look like. This isn't a sales pitch — it's an open engineering direction with a clear next chapter.
Talk to me about this direction