NLP Engineer & Researcher

Dom Marhoefer

Hi! I'm Dom, a Natural Language Processing Master's Student at UC Santa Cruz.

I'm fascinated by computational neurolinguistics (in both biological and artificial systems), and I build pipelines and infrastructure to study the mathematics behind neural communication, linguistic/semantic representations, and why you're able to read this sentence!

Experience

Uniphore
Student Research Engineer, NLP Capstone (six-person team), Palo Alto, CA
May 2026 — Present
Qwen3-OmniSFTRLVR
- Designing a dual-lane benchmark and post-training framework for speech-native tool-calling voice agents, extending an internal multi-agent dialogue simulator with speech perturbation and structured reasoning over noisy, disfluent audio.
- Building the dataset-generation pipeline: dialog simulation, text perturbation (fillers, self-corrections), TTS synthesis (Qwen3-TTS / CosyVoice-3), acoustic noise injection (MUSAN / FSD50K), and automated validation.
- Developing a two-stage post-training pipeline (SFT + RLVR) on Qwen3-Omni to teach direct audio-to-tool-call reasoning, targeting accuracy, JSON validity, and robustness to acoustic degradation.
Intelligent Generation, LLC
Machine Learning / Quant Intern, Remote
May 2026 — Present
LightGBMTFTMPCCVaR
- Built an end-to-end pipeline forecasting day-ahead vs. real-time price spreads across 150+ PJM pricing nodes, combining LightGBM and probabilistic Temporal Fusion Transformer models with calibrated uncertainty estimates.
- Designing a battery-fleet dispatch and hedging framework using model predictive control and portfolio risk optimization (mean-variance, CVaR), with battery degradation costs priced into dispatch decisions.
- Automated daily bid generation, settlement, and simulated P&L tracking via a scheduler pulling market and weather data from public APIs, surfaced through an internal dashboard for market data, bids, and model status.
Camp Kesem
Data Operations Intern, Remote
Jun — Dec 2024
Pandasscikit-learnNLTKVADERLDA
- Applied Latent Dirichlet Allocation (LDA) to decompose 828 survey responses into latent thematic distributions, identifying key drivers of positive and negative sentiment.
- Trained Random Forest ensembles to predict participant retention, surfacing mental health (r = 0.21) and safety perceptions (r = 0.17) as leading predictors; performed feature importance ranking and data visualization to validate model decision boundaries.
- Built a custom preprocessing pipeline using NLTK for lemmatization and stop-word filtering, followed by VADER for lexicon-based sentiment polarity scoring on longitudinal survey data.
Koszyn and Company, LLC
R&D Consulting Intern, New York, NY
Jun — Dec 2023
Regular expressionspython-docx
- Architected a regex-based Named Entity Recognition (NER) system to extract structured insights from highly heterogeneous, unstructured donor prospect data.
- Developed python-docx scripts to automate the ingestion and normalization of semi-structured text, reducing manual data entry latency and improving downstream relational database integrity.
Intelligent Generation, LLC
Data Operations & AI Intern, Remote
May — Aug 2023
scikit-learnPandasMeteostat
- Engineered predictive models for PJM COMED hourly load using 10 years of historical data; integrated weather variables via Meteostat to capture seasonal sensitivities.
- Achieved an RMSE within ~0.5 MW of PJM institutional benchmarks by optimizing hyperparameter manifolds via RandomizedSearchCV.
- Developed a tail-weighted evaluation pipeline aligned with Energy Price Thresholds (EPT), prioritizing model accuracy during peak volatility periods to optimize battery arbitrage logic.

Education

UC Santa Cruz, Silicon Valley Extension

M.S. Natural Language Processing

Santa Clara, CA

Expected Dec 2026

Involvement

Sharf Lab·Independent Study
- Benchmarked GPT-5.4 tool selection over SpikeLab's ~90-tool MCP server for spike-train analysis (80.6% selection accuracy, 72.2% end-to-end), tracing all failures to three reproducible causes and proposing MCP tool-design fixes.
- Prototyped a Python wrapper for the 3Brain BioCam MEA system, working with asynchronous ZMQ streaming of 4,096-channel recordings.
Sep 2025 — May 2026

New York University

B.A. Economic Policy & Language and Mind

New York, NY

Jul 2024

Involvement

Camp Kesem·Development Coordinator
- Led multi-year fundraising initiatives raising $50K+ in total, coordinating 50 counselors and supporting 100+ campers across recurring programs.
2020 — 2024
Zeta Beta Tau·Co-President and Philanthropy Chair
- Directed operations for a 70-member chapter, managing budgets, university relations, and executive board decisions.
- Organized chapter fundraising campaigns and events for national and local nonprofit partners.
2022 — 2024
The B+ Foundation·Team Captain
- Ran campus-wide fundraising campaigns for pediatric cancer research, coordinating volunteers and raising $15K+.
2021 — 2022

Selected Work

PyTorchTransformersRoBERTa

UCSC NLP at SemEval-2026 Task 10: Boundary-Aware Span Extraction and RoBERTa Classification for Conspiracy Detection

2026

Given a piece of text, our system flags which words describe a conspiracy narrative's specific markers (Actor, Action, Effect, Evidence, Victim) and separately judges whether the post is a conspiracy theory or critical narrative.

Subtask 1: Conspiracy Marker Extraction

Built a RoBERTa-large span classifier over enumerated candidate spans (up to 32 tokens), with boundary-aware representations combining start/end embeddings, mean-pooled span content, span-width embeddings, and adjacent context tokens.
Used IoU≥0.95 positive labeling with hard-negative mining (IoU 0.50–0.75) and per-role positive-class weighting to sharpen boundary discrimination under severe span imbalance.
Designed containment-based non-maximum suppression and span merging to resolve overlapping and nested predictions across five conspiracy roles (Actor, Action, Effect, Evidence, Victim).
Ranked 7th of participating systems on the official test set (0.2251 macro F1, token-level IoU≥0.5).

Subtask 2: Document-level Conspiracy Classification

Fine-tuned a RoBERTa-large sequence classifier over the [CLS] representation to predict document-level conspiracy stance (Yes / No / Can't tell), trained independently from the span model with no cross-task feature sharing.
Applied label smoothing and a stratified 90/10 train–validation split to stabilize training under class imbalance.
Ranked 12th of participating systems on the official test set (0.7694 weighted F1).

PyTorchTransformersLlama-3.2

The Capacity of Thought: Benchmarking Llama 3.2 in Semantic fMRI Neural Language Decoding and Improving the Huth Encoding-Model Baseline

2026

Direct fMRI-to-language decoder mapping brain activity into a frozen Llama-3.2-1B, with a blind-control study exposing a language-prior illusion in its apparent success

Originated the project and designed fMRIFlamingo: a brain tokenizer, Perceiver Resampler, and gated cross-attention layers mapping fMRI BOLD activity into a frozen Llama-3.2-1B, adapting the OpenFlamingo multimodal architecture to neural decoding.
Achieved 42.86% Top-1 accuracy on a 1-in-100 next-word ranking task, well above chance — then designed a blind-control ablation (zeroed fMRI input) that revealed the apparent success was driven by the frozen language prior, not the neural signal.
Selected and prepared the Huth et al. narrative-listening fMRI dataset, and diagnosed the failure down to the per-window level (near-identical token distributions between held-out and zero-fMRI conditions) to distinguish it from simple posterior collapse.

Let's Connect

Reach out for

-Collaborations in NLP and computational neuroscience
-ML/DL engineering or applied research opportunities
-Open-source projects at the intersection of language and neural systems
-General questions about my work or experience

Learn more about me

Contact

Emaildominicmarhoefer@gmail.com

LinkedInLinkedIn

GitHubGitHub

Based inSanta Clara, CA

Open to remote opportunities and relocating for the right role.

LinkedIn GitHub