Hardening roadmap

Verification systems decay: fraud techniques evolve, document designs change, models age. This page is the standing process that keeps BAQSHI-KYC improving — and the honest queue of what’s not done yet.

The improvement loop

Every source of “something went wrong” feeds one loop:

detect (monitoring / review / incident / report)
   → record (roadmap entry: signal, impact, evidence)
      → tune or build (threshold, prompt guidance, new check)
         → validate (against the calibration set + the new case)
            → ship (versioned, audit-logged config/code change)

Rules for the loop:

• Every reviewer overturn is a data point. When a human approves what the pipeline flagged (or catches what it missed), the reason goes into the queue.
• Every incident closes with a roadmap entry — Runbooks in Incident response end by filing one.
• Keep a calibration set. Known-good pairs + known spoofs, re-run after every threshold or model change. Never tune on vibes.
• Changes ship via git + deploy — thresholds live in code, so every tuning decision has an author, a diff, and a rollback.

Current queue (honest state, priority order)

#	Item	Why	Status
1	Threshold calibration on real documents	Defaults are literature-derived, not population-tested; auto-approve shouldn’t be fully trusted until this runs	Blocks real users
2	Passive anti-spoof model (e.g. Silent-Face) as a second liveness signal	Closes the pre-recorded-victim-video gap in the threat model	Planned
3	AML screening — sanctions/PEP list matching post-approval	Required for AML/CFT reporting entities; the data feeds are commercial, the matching is straightforward	Planned
4	Review queue web UI	Queue is API-only; a small internal page speeds reviews and reduces errors	Planned
5	Per-tenant API keys + usage isolation	Prerequisite to offering this as a service to others	Planned
6	Redis-backed rate limiting	Current limiter is per-instance memory; needed before horizontal scaling	When scaling
7	Master key in KMS/HSM	Removes the platform-env single point from the threat model	When stakes justify
8	NFC chip reading (passports)	Cryptographic document authenticity beats visual inspection entirely	Aspirational

Reporting a weakness

Anyone using or probing this system: report suspected vulnerabilities privately to the operator (GitHub repository → private security advisory, or the operator’s direct contact). Good-faith research against your own test verifications is welcome; don’t probe real users’ data. Reports get a roadmap entry and a response — that’s the deal.