Evidence-Cited Planning Co-pilot
Making repeat manufacturing planning auditable: every value proposed with its evidence, every change from last time explained.
Role
Sole engineer — assigned by the CEO
Context
750+ person PCB manufacturer
Status
In progress
Verified by
Manual proof-of-concept audit
Results
Material substitutions caught · Manual proof-of-concept
Live error caught before the floor · Manual proof-of-concept
The problem
Repeat orders get planned from scratch. A planner sits down and re-derives decisions that someone already made — often the same someone, months earlier, for the same part.
The history exists. It is just not in a form anyone can consult quickly, so the cheapest path is to redo the thinking rather than go find it.
Constraints
- It has to be auditable. A planner must be able to see why every value was chosen, or the output is unusable regardless of accuracy.
- Wrong answers presented confidently are worse than no answers at all — in manufacturing, a plausible-looking mistake propagates to the floor.
- Human-in-the-loop by design: this is a co-pilot, not an autopilot.
How I built it
The system retrieves over historical job data and attempts every field, but the attempt is not the deliverable — the citation is. Each proposed value carries the evidence it came from, so the planner is reading an argument rather than accepting an assertion.
On top of that sits delta reporting: what changed relative to the prior build of this part, and the stated reason for each change. That inverts the job. Instead of re-deriving decisions from nothing, the planner audits a documented set of proposals, which is both faster and a better use of the expertise they have.
It is in progress. The proof of concept was run manually, and it has already earned its keep: it caught 5 material substitutions and one live error before it reached the floor.
What I’d do differently
Too early to say. The open question right now is how much autonomy is appropriate — every step toward automatic is a step away from the auditability that makes it trustworthy.
Stack
Python · RAG over historical job data · evidence citation · delta reporting · human-in-the-loop design