Darrell S. Best Jr.

I build AI agents and the models behind them.

Senior AI Research Engineer II at USC ISI. I build agent systems that handle real multi-step work, and I fine-tune, edit, and quantize the open-weight models they run on, published on Hugging Face with the eval numbers attached.

View my models See the agent work

USC ISI, Washington, D.C.

Models on Hugging Face: 42
Downloads, last 30 days: 24K
Years shipping applied AI: 7+
Years building agents: 3+

Agent systems

Agents that do the actual work

I've been building agents for 3+ years: systems that run long tool loops against messy real environments, remember what went wrong, and back off safely when something breaks.

Software debloating

Agents that build, then shrink, unfamiliar codebases

Debloating only works on code that builds, and every cut risks breaking it. These agents pull source from git, work out the build requirements, containerize the project in Docker, and drive the build loop themselves. Once it builds, they call debloating tools with increasing aggressiveness, roll back to the last working configuration the moment something breaks, and write up what was removed and why.

Related: An Execution Soundness and Security Benchmark for Java Debloating Tools, accepted poster at ACSAC 2026.

01Clone the repo
02Infer the build
03Containerize
04Build, fix, repeat
05Debloat harder
06Roll back on break
07Report

Agent memory

ECHO

A memory tool I built so agents keep a record of failed builds and fixes. The next run starts from what already went wrong instead of repeating it.

Healthcare

Clinical coding agents

Agents that read unstructured clinical notes, reason over candidate ICD codes, and produce the mapping, cutting down the manual review that dominates medical coding.

Financial crime

Anti-money-laundering agents

Agentic approaches to spotting laundering patterns across transaction graphs, with output that analysts can read and act on.

Open weights

Models on Hugging Face

Every release ships with its numbers: refusal rate, KL divergence against the base model, reasoning and tool-call checks, plus ready-to-run GGUF, FP8, and NVFP4 builds.

Most downloaded 9 repos13K23

Qwen-Image-2.1 Prompt Enhancers

Refusal-ablated versions of the text-to-image and image-edit prompt rewriters for Qwen-Image-2.1, a fine-tuned Qwen3.5-VL 9B. Each one was picked from a 1,000-trial Pareto search, checked through a real editing app's prompt-enhancer path, and repacked as single-file ComfyUI builds.

9/100edit refusals (was 100)

0.0368KL, image edit

0.0130KL, text-to-image

T2II2IComfyUIGGUFFP8NVFP4Text encoder

Vision-language 4 repos4.8K

Qwen3.8-27B Heretic

A 31-trial parameter search landed at 0/100 refusals with less drift than the prior public release. Vision, thinking control, and MTP weights intact.

0/100refusals

0.0465KL divergence

bf16GGUFFP8NVFP4

Small and fast 20 repos4.3K3

Qwen3.5 Heretic series

Edge-friendly 0.6B to 9B models, each evaluated after export and published with quantized builds for llama.cpp and vLLM. The 9B lands at 5/100 refusals.

0.6B0.8B2B4B9B

27B writing model 9 repos1.7K2

Hemmingway-1 Heretic

A Qwen3.8 27B writing fine-tune with refusals removed by full-weight ARA. Reasoning held at 40/40 on GSM8K, the vision tower is restored, and LoRA tuning still works.

0/100refusals (was 98)

0.083KL divergence

bf16Text-onlyGGUFFP8NVFP4

Browse all 42 models

Live stats from the Hugging Face Hub.

Model work

Custom models, fine-tuning runs, and the rest of the pipeline

Each line below comes from a released model or shipped system, with the number that backs it up.

Fine-tuning

LoRA, QLoRA, DPO, full fine-tunes

As a fine-tunability check, Hemmingway-1 Heretic got a rank-16 LoRA on persona and tool-call chats. The adapter merged back cleanly and the model answered unseen prompts with well-formed tool calls.

0.96 → 0.004training loss
Model editing

Arbitrary-Rank Ablation

ARA rewrites attention and MLP output projections directly, no retraining. On Qwen3.8-27B, a 31-trial search reached 0/100 refusals with less drift than the previous public release.

0.0465KL vs. base model
Custom models

Weight grafts and architecture surgery

Hemmingway-1 shipped as text-only weights. I grafted it back into the stock Qwen3.8 vision-language layout, restored the multi-token-prediction head, and checked the result against the source.

1:1logits vs. source
Quantization

GGUF, FP8, NVFP4

Releases get llama.cpp builds from Q4_K_M to BF16, plus FP8 and NVFP4 for vLLM on Blackwell. Each build is re-tested for drift, refusals, and tool calls after conversion.

27 GB27B model in NVFP4
Training at scale

DeepSpeed and federated learning

Multi-GPU fine-tunes of multilingual transformers with DeepSpeed, and federated training with Flower for data that can't leave its owner, using aggregation that holds up when one participant misbehaves.

7+ yrsshipping applied ML

About

Background

I'm a Senior AI Research Engineer II at USC's Information Sciences Institute, based in Washington, D.C. Over seven years I've worked on computer vision for defense sensors, multilingual LLMs, federated learning, clinical NLP, and, for the last three years, agents.

I also release open-weight models on Hugging Face, write on the blog, and I'm finishing an MS in Computer Science at USC.

Where I work

Education

Degrees

MS in Computer Science

2024 to present

University of Southern California, Viterbi School of Engineering. Focus: Data Science.

Applied NLP, machine learning, data mining, information retrieval, algorithms, and database systems.

Graduate coursework

CSCI 544, Applied NLP: transformer language models, fine-tuning, and LLM pipelines.
DSCI 552, ML for Data Science: supervised and unsupervised ML, recommenders.
DSCI 553, Data Mining: frequent patterns, LSH, clustering, link analysis, streaming.
CSCI 561, Foundations of AI: search, CSPs, probabilistic reasoning, planning, game-playing agents.
CSCI 567, Machine Learning: kernel methods, ensembles, deep learning.
CSCI 570, Analysis of Algorithms: dynamic programming, graph algorithms, NP-completeness.
CSCI 572, Information Retrieval: crawling, indexing, ranking, neural retrieval.
CSCI 585, Database Systems: relational and NoSQL, query optimization, distributed data.

BS in Computer Science

2012 to 2017

Clemson University, School of Computing. Minor: Philosophy.

Systems programming, algorithms, and software engineering, plus graduate-level electives in HCI and eye tracking.

Undergraduate coursework

CPSC 1010 / 1020: programming in C and C++.
CPSC 2120: algorithms and data structures.
CPSC 2150: software development foundations.
CPSC 2310: computer organization.
CPSC 3220: operating systems.
CPSC 3500: foundations of computer science.
CPSC 3520: programming systems.
CPSC 3600: networks and network programming.
CPSC 3720: software engineering.
CPSC 4620: computer graphics.
CPSC 4140 / 6140: human-computer interaction (graduate level, Prof. Andrew Duchowski).
CPSC 4120 / 6120: eye tracking methodology (graduate level, Prof. Andrew Duchowski), which led to my ETRA '16 paper.

Publications

Papers and posters

2026

An Execution Soundness and Security Benchmark for Java Debloating Tools

Accepted poster, IEEE ACSAC 2026, the Annual Computer Security Applications Conference, December 7 to 11, Los Angeles.

Accepted
2016

A Rotary Dial for Gaze-based PIN Entry

Best, D. S. and Duchowski, A. T. Proceedings of the Ninth Biennial ACM Symposium on Eye Tracking Research & Applications (ETRA '16), pp. 69 to 76. ACM.

ETRA '16

Google Scholar profile

Toolkit

Stack

Training

PyTorchTransformersPEFT / LoRATRLDeepSpeedLightningJAXFlower

Serving and quantization

vLLMllama.cppGGUFFP8NVFP4OllamaComfyUI

Engineering

PythonC++JavaScriptSQLDockerLinuxCI/CD

Contact

Get in touch

Hiring, collaborating, or curious about one of the models? Email is fastest.