Allen Gary Grimm

Serving as the technical anchor of a flat, collaborative team, I continuously architected new components of a production ML pipeline, each fitting into a growing system I was increasingly the connective tissue of. The pipeline grew to include 10+ external data integrations, processed into nearly 1,000 model-ready features; weekly model retrains with a four-tier versioning system; a human-gated approval process before each model release; and a lead publication framework supporting arbitrary models with configurable thresholds. Over four years, we grew from under 5% to over 85% recall across the full upstream private company space, approaching the theoretical upper limit of what the data could support, and covering the near-entirety of the investment universe relevant to a growth-stage VC firm.

Identified an exponentially scaling recursive join pattern powering our multi-tenant permission system and flagged it as a projected failure at scale. Built a proof-of-concept graph query in Neo4j demonstrating orders-of-magnitude performance gains with linear scalability, but couldn't get organizational traction until the system failed as projected. Led the engineering team through the recovery sprint using my POC as the foundation, migrating the production permission system to the graph-native architecture.

Achieved 80%+ precision on website spam detection and high-quality content identification across a full internet scrape — starting from single-digit precision — by deliberately constraining the model to logistic and linear regression and placing the full burden of performance on feature quality. Engineered rich document features using Word2Vec aggregated into varying Doc2Vec shapes to capture content and integrity signals. The constraint was intentional: rather than chasing model complexity, we proved that disciplined feature engineering over simple models could outperform sophisticated models built on weak features, while keeping minimal compute requirements and enabling reusable feature discoveries across the organization's broader work. It permanently changed how I approach data science problems.

Built a cross-platform identity resolution system merging available user identifiers into a unified identifier with higher fidelity and cohesion across apps and platforms. Publishers supplied initial links between internal IDs and device identifiers, seeding an iterative MapReduce process that derived connected components across all publishers. Careful filtering of edge cases produced a clean unified identifier — the first instance of a canonicalization pattern that became a recurring thread across my career, reappearing at Doing Inc, the Personal Finance Genome Project, and BOND Capital.

Engineered a standalone service to deduplicate and canonicalize events scraped across multiple hosting platforms, where hosts commonly posted the same event on Meetup, Eventbrite, and others simultaneously. Designed complementary distance functions across event attributes to handle sparse and inconsistent records, constructed a matching graph filtered to connected components as candidate duplicate sets, and applied calibrated matching algorithms guided by curated edge cases to resolve sub-graphs into distinct canonical events. Formalized all possible identity mutation rules to support dynamic updates as new scrape data arrived, then delivered the whole system as a regularly-scheduled back-end service built with test-driven development.