Michael Humphrey

Served as one of three primary architects and engineering lead across two cross-functional teams to design and deliver nCino's International Commercial Onboarding & Loan Origination platform — reducing commercial bank customer onboarding from weeks to hours. Live in production with Yorkshire Building Society and one additional UK financial institution, with active international expansion underway via a Dun & Bradstreet integration. Incorporated FullCircl's KYB and eKYC solutions to meet international regulatory compliance requirements.

Contributed as API architect on nCino's greenfield Omnichannel loan origination platform — a unified solution consolidating previously fragmented products across consumer, small business, commercial, and mortgage banking into a single system. Designed RESTful API endpoints for headless deposit account opening. The platform now serves approximately 10% of nCino's client base.

As team lead alongside one engineer and a team of AWS Solutions Architects, architected nCino's Benchmarking & Analytics platform — replacing Domo — by building a multi-tenant QuickSight embedding infrastructure that surfaces loan performance, portfolio risk, and origination lifecycle dashboards within nCino's Salesforce Managed Package. Engineered row-level security ensuring complete data segregation between financial institutions sharing a single dashboard asset.

As sole engineer, designed a fully automated GitOps-style CI/CD pipeline enabling business analysts to promote QuickSight dashboards from development to production with zero engineering involvement — reducing deployment from a multi-hour manual process to ~5 minutes. Automatically validates Row-Level Security guardrails at promotion time, rejects non-compliant dashboards with owner notifications, and versions all dashboard assets in source control.

Architected the core REST API suite for MTRNS, nCino's proprietary banking core integration mapping framework, enabling engineers to centrally create, version, and reuse XPath/XQuery data transformation rules across all financial institution tenants. Replaced bespoke per-client implementations with a reusable multi-tenant repository that standardizes core banking integrations (e.g., Jack Henry jXchange) while preserving per-FI configurability.

Served as lead engineer on a team of 3 building nCino's native Salesforce API orchestration framework — an Airflow-inspired workflow engine built entirely on Salesforce platform technologies. Overcomes fundamental Salesforce governor limit constraints using Queueable Apex, custom Business Activity sObjects, Platform Events, and Flows to orchestrate complex multi-step API workflows. Now powers all nCino integrations including Alloy Fraud Detection and KYC compliance workflows.

Sole engineer on an automated video and image content labeling system for a confidential client, integrating a fine-tuned multimodal AI model (Qwen2-VL) with a FastAPI service to enable automated semantic tagging of media files. The system labels a 10-minute video in 1–2 minutes, generating searchable metadata (scene descriptions, subjects, actions) to power content search and discovery. Deployed FastAPI on AWS ECS Fargate with GPU inference hosted on RunPod and async job processing via SQS. Shipped to production as a core studio feature.

As sole engineer, built a CLI tool and custom YAML parser extending go-swagger to auto-generate Swagger documentation for hundreds of RESTful APIs at build time with configurable public/private endpoint visibility. Enabled developer partners including Whole Foods to self-serve integrations for the first time, directly supporting the launch of FoodLogiQ's formal partner integrations program.

Drove a one-month initiative that elevated FoodLogiQ's integration test coverage from effectively 0% to over 90% across hundreds of API endpoints using swagger-test-templates. Caught real production bugs during the process and established a testing pattern that became a permanent fixture in the API development workflow.

As sole engineer, designed and implemented a fault-tolerant batch product upload system enabling clients to sync millions of products. Engineered SNS/SQS fanout for parallel processing with per-record retry logic and dead-letter queues, ensuring individual product failures never abort a full upload. Enabled Whole Foods to perform full catalog syncs from Glye into FoodLogiQ for the first time.

As sole engineer, designed FoodLogiQ's webhook notification system, eliminating customer polling (some polling every 30 seconds) by delivering near-real-time supply chain event notifications across the entire customer base. Built with at-least-once delivery guarantees, automatic retry on failure, and DLQ handling. Adopted by all customers upon release.

Implemented automated role mapping for SSO, eliminating manual role assignment for newly onboarded users. Roles from users' identity providers — including Azure AD and SAML/OIDC providers via Auth0 — are automatically translated into FoodLogiQ platform roles at onboarding time, meeting enterprise access control requirements.

Served as architect and lead engineer on a 4-month ground-up rebuild of FoodLogiQ's supply-chain traceability platform, replacing a non-scalable in-memory Java queue with a production-grade event streaming architecture. Ingests MongoDB change streams into Confluent Kafka, processes farm-to-shelf events, and loads transformed data into Snowflake for analytics. Designed for ~100 million supply chain events per year. Live in production today.

Sole author of a Python rules-based classification module serving as the first-pass stage in Metabolon's automated metabolomics analysis pipeline. Applies decision-tree classification to HPLC-MS/MS retention time data to identify 30–40 small molecules per study with near-100% accuracy, routing clear-cut identifications away from the computationally expensive multilinear ridge regression stage and reducing analysis time from up to a week to near-fully-automated. Contributed to the patented machine learning workflow (US11387087B2).

Extended Metabolon's legacy data curation application with an Expected Retention Time (RT) visualization derived from the multilinear ridge regression model, displaying expected, actual, and reference RTs side-by-side and significantly accelerating the study review process for scientists.

Built CI/CD pipelines for the Data Science team's AI/ML projects using Azure DevOps, replacing a manual script-execution workflow with reproducible, versioned deployments across all data science Python packages.