The system uses a decision tree diagnostic engine to classify and route incoming tickets, integrates with Jira to close the loop on resolution, and runs with exponential retry logic across Databricks and EC2 to stay resilient when compute or connections fail mid-job. This automated 20% of daily ticket volume and cut resolution time by 80%, work that previously depended entirely on manual triage.

Built LeoLabs' first automated testing suite for satellite tracking systems, replicating production conditions on live data. A Databricks-orchestrated job spins up compute on demand and auto-generates structured Jira initiatives, epics, and tickets directly from test failures, replacing a process that had relied entirely on manual PR review to catch regressions before deployment across an 8-person team.

Retargeting campaigns were reaching only a fraction of the real user base; engaged app users simply weren't in HubSpot. Built an event-driven sync between the application and HubSpot Contacts, with deduplication logic to prevent conflicting records across data sources, and exposed usage properties so marketing could segment audiences by real behavior. This grew the matchable retargeting audience by 62%.

Built a RAG system for student exam prep, but the interesting engineering problem was evaluation, not retrieval or generation. Designed the pipeline around retrieval-confidence scoring and citation gating, then built an adversarial test set using unrelated MCAT sections as distractor content to measure whether the system knew when to abstain rather than hallucinate. It achieved 82% correctness on answerable queries and full abstention on every adversarial input, driving adoption by roughly 50% of users during peak test periods.

Architected the REST APIs and Postgres schemas behind cohort reporting and risk segmentation from scratch. Once in production, table loads slowed significantly under concurrent access; diagnosed the bottleneck as expensive join and count queries on unindexed columns, then added targeted indexes and a Redis caching layer storing precomputed cohort profiles, cutting page-load latency by 35%.

Built the data pipeline behind the Cohort Risk dashboard, aggregating multi-source engagement signals (login frequency, assignment progression) into predictive risk indicators surfaced directly to instructors. This drove a 16% lift in daily active instructor usage of the tool.

Engineered temporal features from GOES satellite solar flux data (lag windows and flux rate-of-change) and designed time-series cross-validation splits specifically to eliminate lookahead bias, a common and easy-to-miss failure mode when forecasting rare, irregular events like SPEs. This reduced feature importance variance by 11% across folds, giving a more reliable signal to build on.

Built the end-to-end modeling pipeline comparing a LightGBM baseline against a Temporal Convolutional Network, using nested cross-validation for hyperparameter selection. The final ensemble outperformed NASA's existing prototype by 6% on out-of-distribution solar event windows, the evaluation setting that actually matters, since SPEs are exactly the kind of rare event a model needs to generalize to, not memorize.

Consolidated 12 legacy PII endpoints to 2 authorized services via granular IAM access controls, reducing blast radius by 83% and achieving ISO 27001 readiness for the startup.

Engineered a RAG workflow agent that cut compliance-change turnaround time by 80% by detecting government portal updates and drafting citation-backed internal alerts using LangGraph.

Engineered Datadog distributed tracing and automated alerting across server infrastructure, enabling real-time anomaly detection that reduced customer service complaints by 22% during server disruptions; customers now get proactively notified when something's wrong instead of discovering it themselves.

Trained a LightGBM churn prediction model on subscription and usage data, flagging users above a 0.8 churn probability threshold and triggering tiered, automated discount outreach calibrated to risk level. This reduced first-month churn by 20% while preserving overall profitability; tiering mattered because blanket discounts would have given money away to users who were never going to churn in the first place.