Privacy engineering · 2026 · Solo — privacy research, numerical architecture, interaction design, build, verification
Noise Ledger
An explainable privacy-budget instrument that turns repeated queries into an auditable epsilon cap decision.
The problem
A privacy engineer planning repeated analytics needs a release-level answer, not one small-looking query at a time. Basic epsilon addition can be too loose, while an unexplained calculator makes budget decisions hard to review. Noise Ledger composes a bounded plan of unsampled Gaussian mechanisms across Renyi orders, converts the total to epsilon at one explicit delta, and exposes why the selected order and dominant query produced the cap verdict.
Architecture
Key decisions
Compose before converting
Privacy loss is added independently at each Renyi order, then converted to epsilon for the requested delta. The engine selects the lowest candidate instead of adding already-converted query estimates.
Make numerical assumptions inspectable
The exposure plot, selected alpha, per-query shares, basic-composition comparison, cap margin, and three monotonicity receipts stay visible together. A reviewer can see both the answer and the mechanism that produced it.
Bound the accountant honestly
The demo supports repeated unsampled Gaussian mechanisms and fixed orders only. It labels omitted contribution bounding, subsampling amplification, PLD accounting, secure noise generation, real datasets, and legal compliance instead of implying production-library parity.
Gate publication with exact deployment evidence
The private product repository is separate from this portfolio site. CI and Vercel deployment records tie the pushed main-branch SHA to the Ready production build; production behavior, rendered English, focus, reduced motion, responsive screenshots, and zero browser errors passed independently. The final build retained a verified 34/35 hiring signal.