Rishabh Maheshwari

Every RAG demo assumes everyone can read everything. Real companies don’t. So the permission filter lives inside the index query here rather than running afterwards on the results, which means content you aren’t cleared for never reaches the app, the prompt or the logs. It never leaves the index at all.

Evidence

evaluation/results.json — 30 questions, 10 documents, 51 chunks. This harness is how I picked the retriever, and Okapi BM25 won by roughly a third more MRR. It also keeps relevance scores identical no matter who’s asking.

A four-month-old account and a four-year-old one churn for completely different reasons, and at four times the rate: 54.9% in the first four months against 14.0% after two years. Train a single model on both and it spends most of its capacity learning tenure. So I didn’t. One model per lifecycle stage instead.

Evidence

artifacts/metrics.json — the dataset is pinned by SHA-256, so two runs give byte-identical output. CI retrains on every push. Drift past 0.02 and the build fails. PR-AUC and Brier sit next to ROC-AUC, because on a problem this imbalanced those are the numbers that actually mean something.

You describe a business problem in English. It comes back solved, with the reasoning shown. The trick is that the language model never gets near the solver — it only emits a typed spec, and everything downstream of that is ordinary code I can unit-test line by line.

Evidence

bench/results.json — ground truth from brute-force enumeration in plain Python. No Pyomo. No HiGHS. Nothing that could share a bug with the code being tested, which is the only reason 5/5 exact is worth saying out loud.