The complaint-intelligence layer that shows its work.

About this project

What Clarion is

Clarion turns messy, free-text complaints into a controlled taxonomy your team can actually trust — because every classification either points to the exact sentence that justifies it, or it doesn't get made. Built on the public CFPB Consumer Complaint Database as a live, working proof: real ingestion, real classification, real human review, running in production today at signal.prateekmulye.dev.

What we offer

Evidence-grounded classification

Every category we assign is backed by a verbatim quote from the complaint. If the evidence doesn't hold up under a strict, deterministic check, we don't guess — we route it to a human instead.

A human review queue that closes the loop

Low-confidence and contradictory cases go to a keyboard-driven triage queue. Every accept or correction is transactional, idempotent, and — for corrections — requires a written reason, which also becomes a new training example for our evaluation suite.

An audit trail you can actually stand behind

Every classification, every review decision, and every model/prompt version is written to an append-only log with no update or delete path, enforced at the database level.

Bring your own model, bring your own data source

Clarion's classification engine is written against swappable adapters — plug in your own LLM provider and your own complaint feed without touching pipeline code. ROADMAP file-drop / SAP QM source adapters are specced, not yet shipped — today's only live source adapter is the CFPB API

Trend intelligence with honest denominators

See how failure modes move over time, compare your data against a live national baseline, and know immediately when that comparison isn't representative — when the national baseline is stale, the page says so before it shows you a number.

Company names, resolved the honest way

Bank names change — mergers, rebrands, charter conversions. Clarion cross-references FDIC's own free public institution records to link a bank's historical names to its current one, and shows exactly which names came from FDIC versus Clarion's own spelling normalization. We never guess: an ambiguous match is left unmerged rather than silently combined. ROADMAP Exact/normalized matches only today — fuzzy or multi-hop merger-chain matching (e.g. bank A renamed to B, B later absorbed into C) is a deliberately deferred, human-reviewed upgrade, not yet built.

How it works

For developers

{
  taxonomy { code label description }
  trends(dimension: FAILURE_MODE, interval: MONTH) {
    periodStart label count
  }
  complaint(id: "1") {
    receivedOn
    classification {
      taxonomyLabel
      status
      evidence { quote }
    }
  }
}

Security

What protects your data

Clarion is a server-rendered analytics product. This is the honest posture for an app of that shape — each item below is a property the system holds today, not a marketing aspiration.

In transit

TLS 1.3, terminated by Caddy with auto-renewed certificates. Port 80 serves only a redirect to HTTPS and the certificate-renewal handshake; no content is ever served in plaintext.

At rest, on disk

The Oracle Cloud volumes holding the database and event log are AES-256 encrypted at rest by default — there is no unencrypted option. Keys are cloud-managed today.

Passwords

Never stored. Only a bcrypt hash (per-password salt, work factor) is kept; the plaintext is discarded the moment it's hashed.

Session & reset tokens

Opaque random values — only their hash lives in the database, so a database or backup disclosure can't be replayed into a live session or a working reset link. Sessions are revocable and expire on a hard 14-day ceiling; reset links are single-use and expire in 20 minutes.

Abuse resistance

Login and password-reset endpoints are rate-limited; authentication responses are constant-time and generic, so they never reveal whether an account exists. The review console sits behind both app auth and a second HTTP-auth layer.

Accountability

Every reviewer decision is written to an append-only log, enforced by a database trigger with no update or delete path, bound to the authenticated identity — corrections capture a written reason and nothing is ever silently overwritten.

Data minimization

CFPB personal information is scrubbed before storage and before any model call; embeddings derive only from the scrubbed text.