@laminesn · AI Scientist · ML Systems · Evaluation, LLMs & MLOps
I build machine learning systems and then find out where they break. Senior AI Scientist at Kegman, previously sole engineer on two production AI platforms at BWX Technologies, and a graduate researcher working on when machine learning evaluation metrics actually carry information.
I'm an AI scientist working where research methods meet production systems. At BWX Technologies, I was the sole engineer on two platforms that shipped: a secure meeting summary system in Azure GCC High used company-wide, and an enterprise spend intelligence pipeline that consolidated four source systems and surfaced a long-standing reporting error.
My M.S. thesis asks when machine unlearning verification metrics actually carry information. I care about systems that don't just demo well, but hold up in production: observable, reproducible, and honest about their limits.
A disciplined toolkit prioritizing correctness, observability, and shipping speed.
ACCEPTED · NEURIPS 2026 WSAn operating-range analysis of an oracle-free membership-inference metric, grounded by retrain and provably memorized references.
ImpactShowed that a published 0.50 null aligns with a memorized member, not erased data; recovered both reference points from held-out data with two forward passes.
SHIPPEDA six-stage Databricks pipeline consolidating four enterprise sources into a canonical schema, then enriching them with rules, a batched LLM, and purchase-order sibling propagation.
ImpactLifted classification from 25.5% to 96.2% across 80K+ transactions, enriched 99.4% of rows, corrected reported spend by 48%, and generated 417 dashboards in 12 seconds.
SHIPPEDAn LLM summarization platform built inside Azure GCC High, from architecture through company-wide rollout, as sole engineer and product owner.
ImpactDeployed to 6,000 employees after a staged security pilot with no data leakage, with human review intentionally built into evaluation and approval.
OPEN SOURCEThree containerized agent tasks spanning log ETL, a Panda3D gridworld, and a Whisper plus FFmpeg video workflow, each with an instruction spec, reference solution, verifier, and rubric.
ImpactVerifiers recompute expected answers independently from source data instead of merely checking that an artifact exists.
RESEARCHReusable PyTorch module with a custom entropy loss and HSIC-based regularizer for analyzing and improving representation learning in convolutional networks.
ImpactPublished in Entropy (MDPI, 2026) — improves convergence and accuracy across CNN architectures.
RESEARCHEnd-to-end pipeline aligning Gazepoint eye tracking with video to train a ResNet-18 encoder-decoder on 25,511 gaze-labeled frames.
ImpactReduced localization error 27.2% against an always-center baseline; 45.4% of predictions landed within 100px on held-out sequences.
ARCHIVEDMulti-agent LLM assistant built on LangChain and FAISS, with a dual-LLM security pattern isolating untrusted content from the privileged model.
ImpactAn MLflow, Dagster, and DVC stack with PEFT-based personalization cut model-release time by 20%.
OPEN SOURCEGeospatial pipeline identifying schools, hospitals, and fire stations exposed to FEMA-designated flood zones using live FEMA and OpenStreetMap data.
ImpactRanks critical infrastructure by flood proximity across Houston, enabling data-driven emergency planning and resource prioritization.
Roles building ML infrastructure, research prototypes, and the platforms beneath them.
Testing whether practical forgetting metrics remain meaningful without a retrained oracle.
Calibrated metrics, reference points, experiment design, and reproducible claims.
How deep networks organize information across layers, channels, and training time.
Latent mutual information and HSIC-based regularization for CNN training dynamics.
Multi-agent LLM orchestration, tool use, and independently verified autonomous workflows.
Reproducible training, evaluation, and deployment of LLM-backed applications.
Attention modeling, gaze supervision, and CNN interpretability on real-world video.
Dagster, MLflow, DVC, and CI/CD for honest, observable model delivery.
Local-first inference and dual-LLM security patterns for sensitive RAG workflows.
Work at the intersection of information theory and deep representation learning.
Studies neural network learning through an information-theoretic lens, measuring mutual information between inputs, hidden layers, and labels across CNN architectures. I contributed the HSIC-based regularizer and the experimental framework.
Accepted · Sole author
Supplies the two reference points an oracle-free unlearning metric lacks: a model retrained without the data and a record the model provably memorized. Shows that the published null aligns with memorization, while both references can be recovered without retraining.
Open to research collaborations and to AI/ML research and engineering conversations.
Recruiters, founders, engineers, researchers, collaborators — if you're working on something interesting at the intersection of software and AI, I'd love to hear from you.
lamineszn@gmail.com