All work

Fintech adjacent

IceVault

How it works

Finance-adjacent spend surface that pairs heuristics and models without bypassing treasury controls.

What we designed

Teams bolt ML onto ledgers quickly, skip lineage, then cannot explain anomalies to finance partners or auditors.

IceVault tags spend with dual confidence scores, attaches policy references, emits alerts with replayable payloads, and never mutates authoritative ledger cores.

Evidence bundles

Each classification ships with supporting transactions, heuristic codes, embedding distances, and analyst notes zipped for audit pulls.

Model + rules sandwich

Deterministic freezes handle regulated categories while models prioritize merchant noise—ordering is contractual, never accidental.

Alert choreography

Notifications stage through severity ladders; treasury webhooks ACK only once downstream systems mirror state.

Flow
  1. 01

    Sanitized ingestion

    Connectors hydrate read models from accounting APIs with checksum validation and dead-letter trenches for poison rows.

  2. 02

    Dual-path classification

    Heuristics short-circuit known merchants; residuals enter embedding + calibrated classifier lanes with calibrated thresholds.

  3. 03

    Anomaly choreography

    Spikes fan into parallel investigations tracked as cases—each step references policy IDs and SLA timers.

  4. 04

    Controlled feedback loop

    Analyst corrections write to golden sets; promotion gates evaluate drift dashboards before nightly batch retrains.

Under the hood

Batch + streaming ingestion, Postgres + immutable append-only ledger mirror, KMS-wrapped secrets, PCI-scoped subnets when required.

Stack highlights

Python workersPostgresFeature storeKMSAirflow-lite

Illustrative composite based on recurring studio patterns—not a testimonial tied to any single deployed client.