AHTBE 2026 — 2nd International Conference on Advancement in Healthcare Technology and Biomedical Engineering · August 21, 2026
Hybrid NLP at Scale: Optimizing Clinical De-identification for High-Throughput FedRAMP Lakes
De-identifying unstructured clinical text at petabyte scale runs into a hard tradeoff: Transformer models like ClinicalBERT catch context-dependent PHI but are too slow for high-volume ingestion, while fast RegEx systems miss it and leak privacy. This paper introduces a hybrid triage architecture that uses lightweight heuristic scanning to process 90% of low-entropy clinical text at line speed, routing only ambiguous or high-context segments to a full Transformer model. Benchmarked in a production FedRAMP High environment, the hybrid approach matches full-model accuracy (F1 0.98) while cutting compute costs by 40% and ingestion latency by 60%, offering a practical blueprint for balancing privacy compliance with research velocity.
PyCon Korea 2026 (Deep-Dive Workshop) · August 17, 2026
Go-Live at Hour Three: An Observability War Room
A full-day, hands-on workshop simulating a live production incident: teams build a real observability stack — structured logging, metrics, alerting, contract tests, and trace correlation — against a working-but-blind data pipeline in the hours before a new partner integration goes live. When the partner goes live at hour three, staged waves of realistic failures (schema drift, unsized volume, poison records, silently wrong data) test whatever instrumentation the team chose to build, with missing signals added live as blind spots surface. Requires only Python and a laptop — no monitoring or security background needed — and ends with a retro on which signals mattered and where the team was still blind.
Chapter 15 in Generative AI Risks and Benefits within Human-Machine Teams (Elsevier) · August 1, 2026
Synergistic Pedagogy: Integrating AI Collaborators into Data Science Education
Book chapter (corresponding author) on integrating AI collaborators into data science education. Forthcoming — publication pending.
USENIX PEPR '26 · June 1, 2026
The Disposable Identity: Eliminating Non-Human Identity Risk in Federal Healthcare Pipelines
Non-human identity — the long-lived, over-privileged service accounts powering automated pipelines — is the fastest-growing and least-examined attack surface in cloud data environments; in federal healthcare systems under FedRAMP High, a single compromised ingestion role can mean bucket-wide access for up to 90 days. This talk presents a production case study of an Identity-Per-Transaction (IPT) pipeline deployed for a federal life sciences agency that issues a unique, cryptographically scoped, ephemeral credential for every file-ingestion event and destroys it milliseconds later, including the operational realities of running it in production — latency, concurrency race conditions, and debugging credentials that no longer exist — plus an honest reckoning with the approach's limits: the identity broker remains a bounded but non-zero root of trust, and eliminating credential-mediated access risk is not the same as solving genomic data anonymization.
Learn more →
PyCon US 2026 · May 16, 2026
Zero Trust in 200ms: Implementing Identity-Per-Transaction with Python and Serverless
Building a serverless data pipeline that satisfies FedRAMP High while handling Protected Health Information is typically a nightmare of encryption management and rigid access controls. This talk dissects a production federal life-sciences pipeline that replaces static service accounts with ephemeral Python logic: an identity broker that mints a unique, cryptographically scoped IAM credential for every file transaction and destroys it milliseconds later, a streaming de-identification layer built on Python generators and Microsoft Presidio to tokenize PII in-memory before it reaches the data lake, and structured logging patterns that produce immutable, audit-ready JSON trails — showing how Python can solve the "Non-Human Identity" crisis in high-stakes environments.
Learn more →
IEEE BigDataSecurity 2026 · May 10, 2026
Zero-Trust Data Engineering: A Reference Architecture for Serverless, FedRAMP-High Healthcare Pipelines
As healthcare research migrates petabyte-scale genomic and clinical data to the cloud, traditional perimeter-based security models violate zero-trust principles and let a single compromised credential expose an entire data lake. This paper proposes a shift from identity-per-user to identity-per-transaction, presenting a serverless zero-trust reference architecture (AWS Lambda/Glue) that uses a just-in-time identity broker and a clean-room, streaming de-identification layer to scope permissions to the lifecycle of a single batch event. Evaluated in a production FedRAMP High environment, the architecture cuts the blast radius of a compromised identity by over 99.9% while preserving linear throughput scalability — proving that rigorous federal compliance doesn't have to come at the cost of cloud agility.
Learn more →
AAAI 2025 Spring Symposium Series — Current and Future Varieties of Human-AI Collaboration · March 31, 2025
AI as a Collaborative Partner in Data Science Education
This talk explores how AI can partner with instructors and students to make data science education more engaging, effective, and equitable — from AI tutoring systems that give real-time, personalized feedback on coding assignments and support students with learning disabilities, to automated grading tools that free instructors for higher-level teaching, to collaborative projects where students and AI agents jointly tackle data analysis and modeling problems — while arguing that these gains must be paired with deliberate attention to bias, fairness, privacy, transparency, and accountability in AI's classroom deployment.
Learn more →