J·H
N° 00 / Maison J·H S/S 2026 — En vigueur

JazmiaHenry

Principal Research Engineer · Foundation Models · RL Systems · Evaluation · Efficient Inference

Enter the atelier
N° 01 / 05

On record.

Fifteen years building production AI systems — most recently the RIGGs platform at Collide, a 120B-parameter mixture-of-experts model for petroleum engineering whose 8B variant outperforms GPT-5 and Claude 4.5 on domain benchmarks at orders of magnitude smaller scale.

Research spans the full reinforcement-learning taxonomy — PPO, SAC, GAIL, UCT-MCTS, DPO, PPO-RLHF, EM-MIA-weighted DPO — large-scale distributed continued pre-training via NeMo + Megatron-Core and DeepSpeed ZeRO-3, multimodal embedding ensembles, agentic evaluation, and AI fairness. Two NeurIPS submissions, six reviews on the Evaluations & Datasets track, one USPTO-published patent.

Maison
Collide · Principal MTS, AI/ML
Founded
Iso AI · 2024
Patent
US20260080257A1
Education
Oxford (DPhil, ABD) · Stanford HAI · Columbia · Tulane
N° 02 / 05

Selected work.

Look 01 — Atelier piece

RIGGs · Collide

2025 — present · Principal MTS, AI/ML · Backed by Mercury Fund

Full ownership of the RIGGs platform — a purpose-built mixture-of-experts LLM for petroleum engineering, trained on Collide's on-premise Spindletop cluster (NVIDIA Blackwell 6000 Max-Q · AMD Threadripper PRO). miniRIGGs — an 8B variant — outperforms GPT 5.1, Claude Sonnet 4.5, and Grok 4 on the SPE petroleum engineering exam at orders of magnitude smaller scale. BigRIGGs — 120B MoE on a 55B+ token domain corpus — ships with a 16-dimensional reward system, simulation environments for retrieval, routing, calculation, and long-context reasoning, and a custom multimodal embedding stack. Sub-100ms response on the MoE at production load via sparse expert routing.

N° 01 — Benchmark

SPE Petroleum Engineering Exam · 40-question subset · May 2026

ModelScoreTime
RIGGs67.5 %15 min
Grok 462.5 %2 hr
Claude Sonnet 4.552.5 %—
GPT 5.14.0 %—

Target accuracy: 75 – 80 % within the coming quarters.

N° 02 — Operational impact

Winn Resources · Texas Railroad Commission filings, W-10 & G-10

  • 95 %+ Processing-time reduction on regulatory filings.
  • 50 wells Filed in 20 minutes — a previously multi-hour manual workflow.
  • 2.5 × Accuracy improvement from domain-tool integration.

N° 03 — Custom embeddings

A specialized retrieval stack for oil & gas — domain tokenizer, vision pairing, and a knowledge graph for grounded reasoning.

  • PE-smol — 397M-parameter dense embeddings with a custom oil-and-gas tokenizer.
  • PE-large — 7B-parameter multimodal embeddings for subsurface and field-engineering documents.
  • Vision pairing — SigLIP + EVA02-CLIP two-encoder ensemble with expert-aware projection heads and a learnable-temperature sigmoid contrastive loss, trained via a four-stage curriculum (contrastive alignment → ontology grounding → hard-negative mining → cross-modal fine-tune).
  • Knowledge graph — Neo4j, 70,071 nodes, 652,447 relationships, integrated for retrieval-augmented training and inference.

In the press

  • NeMo
  • Megatron-Core
  • DeepSpeed ZeRO-3
  • FSDP2
  • vLLM
  • SGLang
  • AWS Optimum Neuron
  • LangGraph
  • Neo4j
  • Qdrant
Look 02 — Founder line

Iso AI

2024 · Founder & Chief AI Officer

RL simulation engine running 10K+ concurrent episodes at sub-100ms. IsoTune — EM-RL fine-tuning combining Expectation-Maximization self-training with PPO-driven decision-making and novel EM-MIA-weighted preference re-weighting. ISOPro — open-source evaluation replacing learned reward models with deterministic verifiers; runs on a MacBook under 8GB RAM.

  • USPTO Patent
  • NeurIPS 2026 (under review)
  • Ray
  • Gymnasium
  • RLlib
Look 03 — Maison Microsoft

Project Bonsai · Microsoft Research

2022 – 2024 · Lead Applied AI Engineer

Co-developed Azure Plato — distributed training platform and RL package that cut training time 80% through architectural optimization. Built production RL agents for Abbott Nutrition, Shell, PepsiCo, BlackRock, and Moody's; 27% portfolio performance improvement on the financial line.

  • SAC
  • GAIL
  • LangChain
  • LangSmith
Look 04 — Heritage

The Motley Fool

2020 – 2022 · Head of Machine Learning

Built the ML organization from zero to seven. 95% churn reduction company-wide. $40M revenue impact. $400M valuation lift.

Look 05 — Provenance

Morgan Stanley

2018 – 2020 · Senior Data Strategist

Long-horizon Monte Carlo simulations for UHNW wealth-management clients and institutional accounts. $2B+ AUM in aggregate.

N° 03 / 05

Archives.

  1. N° 01

    Beyond Static Snapshots: A Grounded Evaluation Framework for Language Models at the Agentic Frontier

    Jazmia Henry · arXiv:2604.17573 · NeurIPS 2026, Evaluations & Datasets Track — under review

  2. N° 02

    Scaling Laws or the Law of Diminishing Returns

    NeurIPS 2023, Black in AI Workshop — Top 5% Paper

  3. N° 03

    Ethics in Action: Comparative Analysis of Philosophical Frameworks in Resource Allocation using AI Agents

    Published paper · Rawlsian, utilitarian, individualist allocators

  4. N° 04

    Combatting Bias in Large Language Models

    Chapter · McGraw-Hill

  5. N° 05

    AI Alignment and Ethics

    Chapter contributor · Bloomsbury Academic

  6. N° 06

    AAVE Corpus

    141,000+ word open-source dataset of African American Vernacular English · Stanford HAI · 21★ on GitHub

  7. N° 07

    Your AI Agent Isn't Broken — It's Poorly Designed

    Jazmia Henry · Medium · June 2025 — A tripartite framework (conceptual · empirical · technical) for designing production AI agents. Read  →

N° 04 / 05

The runway.

N° 05 / 05

Inquiries.

For correspondence regarding research, speaking, or commissions —