I build reliable AI and infrastructure systems for high-stakes operations.

I'm an AI infrastructure and applied AI engineer with experience spanning production financial infrastructure, security and network automation, responsible AI research, and developer tooling.

At LG CNS America, I work with financial institutions on network and security systems across data-center, disaster-recovery, and branch environments. I also build Python/API automation for configuration, monitoring, audit evidence, and operational workflows.

Earlier, I researched differential privacy, reproducibility, and machine learning robustness at NYU's Center for Responsible AI.

A safety boundary for AI-generated infrastructure changes.

ChangeSafe converts AI recommendations into typed, declarative patches and validates them through deterministic policy, transactional simulation, rollback verification, and human approval before they can become executable.

The v0.1 prototype focuses on:

• Schema and state-path validation
• Blast-radius and protected-resource policies
• Management-reachability checks
• Fail-closed approval states
• Inverse-operation rollback testing
• Hash-bound decision receipts

Production execution is intentionally disabled in v0.1.

A local-first context compiler for humans and AI tools.

GotSaeng OS turns scattered Markdown and Obsidian notes into auditable, model-ready context packs with deterministic memory diffs, source-provenance scoring, confidence signals, and contradiction review queues.

• Published CLI: @gotsaeng/cli
• Local-first with no telemetry or cloud dependency
• Monorepo with core compiler, CLI, and Obsidian adapter
• Automated type checking, testing, build, lint, and CI

Co-author of "Epistemic Parity: Reproducibility as an Evaluation Metric for Differential Privacy" (PVLDB).

• VLDB 2023 Evaluation & Benchmark Track Runner-Up
• ACM SIGMOD Research Highlight
• Contributed to SynRD, an open-source framework for reproducibility research using synthetic data

Co-author (with S. Rahman) of "Out of Distribution Performance of State of Art Vision Model" (arXiv:2301.10750) — Computer Vision, NYU.

• Benchmarked 58 state-of-the-art vision models — convolutional, attention-based, hybrid, and MLP families — across five out-of-distribution benchmarks (ImageNet-A, -R, -O, -Sketch, Stylized-ImageNet) under a unified training setup
• Findings challenge the claim that vision transformers are inherently more robust than CNNs: robustness tracked training configuration and shift type, not architecture family
• Source

Co-author (with V. Mavi, S.H. Chi, D. Jacob) of "Vision Model for Temporal Disease Progression of Chest X-ray dataset" — Data Science for Healthcare, NYU.

• Classified disease progression (improving, stable, worsening) from longitudinal chest X-ray pairs on the MS-CXR-T benchmark, using TorchXRayVision DenseNet121 encoders
• Owned the dataset pipeline and evaluation across three modeling approaches
• Source

Additional coursework and research:

• Replicated Concurrency Control and Recovery

Software & AI Python · TypeScript · Flask · REST APIs · PyTorch · Pandas · SQL · Docker · LLM/agent workflows · AI evaluation · Reproducibility

Infrastructure & Security Cisco IOS/IOS-XE · Firepower/FMC · BGP · MPLS · VPN · ACL/NAT · Network automation · Monitoring · Change validation

I'm especially interested in systems where AI moves from recommendations into real operational decisions—and where evaluation, permissions, observability, rollback, security, and human oversight are part of the engineering problem.

LinkedIn · Google Scholar · Email