HOUSTON, TX — DATA & AI
Enterprise-grade by day. Open-source by night.
Data & AI Manager at EY. I design lakehouse platforms on Databricks and Azure, and ship applied AI — RAG, agents, MCP — that turns complex data into business outcomes.
After hours I publish open models and tooling for auditable AI: benchmarks, evals, quantization ladders. Everything measured, nothing vibes-based.
01 / ABOUT
The same instinct runs through both: systems you can trust because you can inspect them.
BY DAY
Leading Data & AI delivery at EY — lakehouse architecture, governed pipelines, and applied-AI systems that have to work for real businesses, at real scale, under real scrutiny.
BY NIGHT
Building the tooling that makes AI auditable — open audit-domain models, tamper-evident RAG trails, eval harnesses, and quantization ladders where every claim ships with a measurement.
// the overlap: I work at an assurance firm — and I build tools
// that make AI systems auditable. Same instinct, different stacks.
02 / EXPERIENCE
Seven-plus years from data pipelines to AI platforms.
EY · Assurance Services
Aug 2024 – Present · Houston, TX
Leading data & AI for EY Assurance — architecting the audit data platform and shipping production LLM systems used by 1,000+ engagement teams.
EY · Assurance Services · Oct 2021 – Jul 2024
EY · Assurance Services · Nov 2019 – Sep 2021
EDUCATION
University of Houston–Clear Lake · 2019
EDUCATION
Mumbai University · 2017
03 / OPEN SOURCE
Open models, benchmarks, and tooling for AI you can actually trust — every release benchmarked before it ships.
THE AUDITABLE-AI STACK
auditlm
Open, locally-runnable language model and benchmark for US GAAP & PCAOB-based external audit.
Python · Fine-tuning · Local LLM
assurancebench
Benchmark for evaluating language models on external-audit work — with a safety and guardrail suite.
Python · Benchmarks · Guardrails
MORE TOOLS
Find where your pandas/polars pipeline silently broke — nulls, dropped rows, dtype drift. No rules to write.
AI-powered search and alerts for the US Federal Register.
Synthetic OCR corpus + CER/WER harness for measuring exactly what quantization costs (MLX + GGUF).
Download counts via pepy.tech (official PyPI BigQuery data), refreshed daily.
13 models · 2 datasets · 4 spaces · refreshes daily
+ 8 more variants, incl. the HY-Embodied-0.5 MLX ladder (4/5/6/8-bit & bf16) — every quant published with a measured eval ladder, not vibes.
DATASETS
stillwarm-bench-results
Benchmark results behind stillwarm — measured cold vs. warm KV-cache timings across local models.
— downloads
stillwarm-kv-cache-artifact
Reproducible KV-cache artifact powering the stillwarm restore demo — load it and skip prompt re-processing.
— downloads
04 / SKILLS
DATA & PLATFORMS
GENERATIVE AI & LLMs
TOOLS, APPS & RESEARCH
05 / PUBLICATIONS
arXiv:2607.09682 — the research behind the open-source AuditWeave project; provenance you can hand to an auditor.
REVIEWSurvey of IoT architectures and protocols for smart home automation.
All publications on Google Scholar →06 / CERTIFICATIONS
07 / WRITING
Overlooked fundamentals and new AI tools I've personally tested — everything from real use, not announcements.
DEV.TO
CER/WER ladders across MLX quants — and why mixed 4-bit beat straight 4-bit.
Jul 2026 · 7 min read
DEV.TO
Tracing silent row loss, nulls, and dtype drift through pandas pipelines.
Jun 2026 · 5 min read
HUGGING FACE
The real cost of cold KV-caches in local workflows — and how stillwarm fixes it.
Jul 2026 · Article
07 / CONTACT
Always up for genuinely hard data problems. If you need AI that can survive an audit — literally — say hi.
Connect on LinkedIn