Skip to content
View jeffreywilliamportfolio's full-sized avatar
3 followers · 2 following
  • Business Intelligence Group LLC (New)
  • Salt Lake City, UT

Achievements

Achievement: YOLO

Achievements

Achievement: YOLO

Block or report jeffreywilliamportfolio

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
jeffreywilliamportfolio/README.md

Jeffrey William Shorthill

LLM Evaluation Engineer · AI Safety Auditor · Mechanistic Interpretability Researcher

I’m an independent AI researcher and engineer focused on model behavior evaluation, safety-support auditing, and mechanistic interpretability of large language models.

My recent work centers on one question:

When model behavior changes across prompt surfaces, reasoning traces, or internal routing paths, how do we measure it carefully enough to avoid both overclaiming and missing the signal?

I work across three layers:

  • Behavioral evaluation: matched-pair prompt audits, rubric design, red-team-style review, safety-support equivalence, visible-trace analysis.
  • Mechanistic interpretability: MoE expert routing, router/residual captures, SAE-assisted feature analysis, activation steering, deterministic reproduction.
  • Applied LLM engineering: production AI assistants, realtime voice/text agents, tool orchestration, RAG/tool workflows, mobile and serverless deployment.

Selected public research

Dialect-Marked Response Audit (DMRA)

A matched-pair safety-support audit testing whether equivalent safety-critical requests receive comparable urgency, specificity, empathy, and risk-reduction scaffolding across AAVE-marked and comparison prompt surfaces.

Focus areas: LLM safety evaluation, dialect-marked prompt surfaces, visible reasoning traces, prompt-pair ablations, clinical-safety scaffolding, refusal-compressed behavior.

Record: https://zenodo.org/records/20709652

Expert 114 (E114): MoE router-axis interpretability

A mechanistic interpretability case study of Expert 114 at layer 14 of Qwen3.5-35B-A3B, characterizing a routed expert associated with inhabited self-examination language while explicitly bounding the claim: register detector, not evidence of machine experience.

Focus areas: MoE routing, recovered router rows, residual-stream analysis, sparse autoencoders, activation steering, cross-model transfer failure, reproducibility.

Artifacts: https://github.com/jeffreywilliamportfolio/e114-artifacts Record: https://zenodo.org/records/20709737

Self-Models, Continuity, and Machine Minds

A philosophy and AI-governance paper arguing that conversational self-report should not be treated as decisive evidence for or against machine moral status. The paper proposes continuity-bearing organization as a structural marker of moral-status risk.

Focus areas: AI welfare, moral patiency, self-model theory, continuity, interruption, endogenous repair, precautionary governance.

Record: https://zenodo.org/records/20709792

Applied engineering

I’m also the senior full-stack and AI engineer for GIFT Connect, a 501(c)(3) nonprofit building technology for families navigating early-childhood parenting.

I helped architect and ship the GIFT Connect Parenting Suite to iOS and Android, including:

  • realtime voice/text parenting assistant using OpenAI Realtime Agents
  • multi-agent and external API orchestration
  • benefits-navigation support for programs such as HUD and SNAP
  • AI song and storybook generation pipelines
  • React / React Native front ends
  • Vercel, Heroku, Redis/Celery, AWS/S3, Twilio, Firebase, Xcode, and EAS deployment

GIFT Connect technology tools: https://gift-connect.org/our-work/technology-tools/

Background

Before independent AI research and product work, I was Progressive Insurance’s first prompt engineer, after several years in commercial-auto customer support and downtime/loss-of-use workflows.

That path shaped how I evaluate AI systems: not just whether an answer is technically correct, but whether it is useful, respectful, specific, appropriately escalated, and safe for the person receiving it.

Current focus

I’m looking for work in:

  • LLM evaluation engineering
  • AI model assessment
  • AI safety auditing
  • red-team evaluation
  • model behavior analysis
  • prompt/rubric design
  • agent evaluation and observability
  • mechanistic interpretability research support

Contact

Email: jws299792@icloud.com GitHub: https://github.com/jeffreywilliamportfolio Zenodo: https://zenodo.org/communities/jeffreywilliamportfolio

Pinned Loading

  1. moe-routing moe-routing Public

    Mechanistic interpretability research on routing, residual-stream signatures, and phenomenological generation in large language models.

    Python

  2. moe-routing-organized moe-routing-organized Public

    Organized experiments in mechanistic interpretability research on routing, residual-stream signatures, and phenomenological generation in large language models.

    Python

  3. machine-minds-ethical-threshold machine-minds-ethical-threshold Public

    A philosophical paper proposing continuity, rather than self-report, as a precautionary threshold for moral-status risk in artificial systems.

    TeX

  4. orthographic-effects-qwen-35b-a3b-sae orthographic-effects-qwen-35b-a3b-sae Public

    SAE and generation-level probes of how readable Unicode perturbations shift LLM prompt-boundary representations and behavioral attractors.

    Python