About the role
Most expense software reads a receipt the way a busy manager does: it finds the total, checks it against the card charge and moves on. Phyto Nex AI reads the folio the way a forensic accountant would. A four-night hotel stay is not one number; it is room rate, resort fee, two minibar charges, a spa visit posted to the room and a parking line that belongs on a different cost center. Our audit engine can only enforce policy on what our vision models can see.
This role sits at the center of that problem. You will own the models and evaluation harness behind line-item extraction across roughly 310,000 receipts a month, spanning thermal paper photos, PDF folios from 2,300+ hotel properties, Japanese and German restaurant checks, and ride-share breakdowns. The work is applied research with a direct line to revenue: every point of recall on a misread alcohol line is money a controller recovers and a conversation an AP manager does not have to have.
What you will do
- Own the line-item extraction models for folios, itemized checks and flight receipts, from labeling strategy through production rollout.
- Build and maintain an evaluation harness with per-merchant-class precision and recall, regression gates and drift alarms that block bad releases.
- Design layout-aware approaches for multi-page folios where charges, taxes and adjustments are spread across pages and reversed on later dates.
- Reduce false positives on policy-relevant fields (alcohol, incidentals, cabin class) where a wrong flag costs customer trust more than a missed one.
- Partner with the policy engine team on confidence calibration so low-certainty extractions route to human review instead of becoming flags.
- Keep inference cost and latency inside budget: p95 under 4.2 seconds per receipt at peak month-end volume.
- Mentor two engineers and write internal design docs other teams can build on.
What you bring
- 6+ years building production ML systems, with at least 3 focused on document understanding, OCR or layout analysis.
- Hands-on experience fine-tuning vision-language or layout models and shipping them behind real SLAs.
- Strong Python and PyTorch; comfortable reading CUDA profiles and fixing data loader bottlenecks yourself.
- A track record of building evaluation sets that reflect messy real-world input, not clean benchmark data.
- Clear written communication. You can explain a precision and recall trade-off to a controller without jargon.
- Comfort working with sensitive financial data under strict access controls, PAN masking and zero-retention rules.
Nice to have
- Experience with hotel folio, invoice or receipt datasets in any language.
- Published work or open-source contributions in document AI.
- Background in active learning or weak supervision for low-label regimes.
- Exposure to accounting concepts like accruals, reversals or VAT reclaim.
Your first 90 days
First 30 days
Ship a fix to the folio reversal parser and own the weekly accuracy review with the audit engine team.
By day 60
Rebuild the alcohol and incidentals evaluation set across 14 merchant classes with merchant-level error budgets.
By day 90
Lead the next extraction model release from offline eval to 100% of production traffic with zero accuracy regressions.