Applied AI Engineer (Remote OK)

Provenance is the AI system of record for the two artifacts every private-markets deal runs on: the Excel financial model and the PowerPoint deck. Think "version control plus audit trail" for deal work, every version preserved, every change shown cell-by-cell with the reason recorded, any prior state restorable, all under MNPI-grade access control. It just closed a pre-seed at a $20M valuation led by Jackson Square Ventures (out of JSV's S26 Launchpad) with Glasswing Ventures, and is raising again toward a ~$50M valuation.

Founded by four young, technical Turkish founders, CEO Tuna Üsküdar (Boston College) plus Arda Dinç, Chris Risio, and Prahaas Nukala, it's selling to PE and investment-banking teams at $50-150k/firm, targeting $1M ARR by February 2027. The record layer works and is sellable today (the founders cover it, plus all frontend and day-to-day product). On top of it sit the two systems that are the actual moat, and the reason for this hire: Attribution (prove where every number came from, a figure on a slide names the exact Excel cell; a hardcoded model assumption names the data-room document that supports it) and the Assistant (a read-only analyst that answers deal questions grounded in server-verified citations, never model confidence).

Provenance competes with Rogo and Hebia but pushes into multi-step agentic workflows. It's a mature production monorepo, 60+ API modules, 25+ data pipelines, eval-gated in CI, on Python/FastAPI, Postgres. pgvector, Gemini-on-Vertex (OpenAI for extraction), with Office add-ins.

Applied-AI engineer who owns Provenance's intelligence layer end to end, the attribution engine and the Assistant, while the founder covers the record layer, the frontend, and day-to-day product. This is deep work in a real system, not greenfield, and it's backend/infra only (no frontend). The person needs to be in seat by mid-October; the six-month plan sequences around them. The mandate is sequenced, not parallel. Week one: two switch-sized Assistant fixes ship behind existing telemetry, turn the grounding verifier back on and raise the throttled tool-round budget.

Month one: finish the deck-attribution leg to demo-grade (it already agrees 120/120 where PowerPoint declares the link). From there: the long pole, rebuild document-attribution recall (stuck at ~0.30 against a 1.0 bar because there's no semantic retrieval, and precision-protecting binary gates delete true positives), converting binary gates into calibrated scores to recover true matches without giving back precision, while growing the Assistant's eval corpus (8 cases today to 150+).

The core mission: take recall from 0.30 to "e0.6 at precision "e0.9 on expanded labeled sets, make "every number names its cell" real, get the Assistant to analyst-grade (reliability "e0.9), mounted inside Excel, with a first human-approved write path. Beyond six months, this grows into ownership of the whole intelligence layer.