Five years at Accenture working on data infrastructure and AI platform integration for a Fortune 100 bank. I like understanding how things break — I read papers, run experiments, and tend to write up what I find. Current side project is figuring out where LLM guardrails actually fail.
Fortune 100 financial-services client · Milwaukee, WI
Go + Kubernetes foundation-model gateway on AWS Bedrock — all org LLM workloads · 5-person team
- Engineered E2E integration test harness — validated Codex + third-party integrations against hard deadlines using Claude Code
- Built enterprise chargeback logging — AI spend attributed by team across all foundation-model workloads, enabling org-wide cost accountability
TSYS + Finicity unification — personal-loans data across disparate systems · Lambda + DynamoDB
- Built GraphQL data-product API across 3 ETL schemas — one interface replacing per-consumer integration work
- Multi-step orchestration across payment processor + external providers — consumers get complete data, no call coordination required
Legacy COBOL account data · S3 → Snowflake
- Airflow pipeline — zero consumer downtime, cutover ran behind existing views throughout
- Unified data ontology across legacy/target — migrated records consistent on new platform
Large-scale migration program · Databricks, Snowflake, Kafka · Scala, Python, Java, SQL on AWS
- 1,000+ dataset migrations (6 pattern types); 250+ streaming with zero downtime
- Replaced shell-script ETL — release velocity 1–10 → 20–25 datasets/cycle
- Config-driven validation across ~1,500 datasets — 3 min → 30 sec/dataset, ~62 hrs eliminated per cycle
- One of 7 engineers selected org-wide for AI coding pilot
- Co-designed KNOWLEDGE/SPEC/PLAN methodology — 66% reduction in per-story delivery time; published as series
Credit-card decisioning system · PostgreSQL-backed distributed pipeline
- Led data integration + schema design for credit-freeze flow across full card-acquisition lifecycle
- Built test-vetted UI for credit-policy modelers — prototyping days → hours
- Found + fixed Kafka consumer test suite blind for months — prevented silent data-correctness failures reaching consumers
- PySpark 2→3 migration of production Step Function ETL — zero downtime
Core financial-services data infrastructure · audit-compliance constraints · multi-platform decisioning
- Built production streaming pipelines; contributed to DMN rules engine + credit-card decisioning data infrastructure on PostgreSQL
- Designed multitenancy schema + data exhaust architecture — business-unit isolation over shared ETL infrastructure
- Drove schema modernization across platforms — aligned legacy/modern contracts for decisioning-system consumers
Algebraic and Computational Limits of LLM Guardrails
Independent research on where language-model guardrails provably fail — regex filters shown blind to modular-position encodings via syntactic-monoid aperiodicity. 5 attack vectors, monoid-extractor audit tool, benchmark harness (ToT+LLM 4.3× over BFS, p<0.001). Solo-authored preprint — not peer-reviewed.
AI / Platform: LLMs, AWS Bedrock, Claude Code, Codex, prompt engineering, spec-driven development, AI platform integration & validation, enterprise cost instrumentation
Languages: Python (Pytest, Pandas, PySpark, Airflow), TypeScript, Java, Scala, SQL
Cloud & Platform: AWS (Bedrock, Lambda, DynamoDB, Step Functions, S3), Kubernetes, Docker, GCP, Terraform, CI/CD
Data: Snowflake, Databricks, Kafka, PostgreSQL, Airflow, PySpark, Spark/Scala, data ontology & schema design
Backend & APIs: REST, GraphQL, OpenAPI, Java/Spring Boot, DMN rules engine
Tools: Claude Code, GitHub, Splunk, Jira, Confluence