Software engineer, AI researcher in training, technical consultant, and educator. I build production AI for government and community programs, and I teach engineers to build them responsibly.

• Production AI for government benefits: document intelligence, multilingual conversational AI, fraud-signal detection, bias testing, and high-availability payments infrastructure for state benefits programs (Michigan, Colorado) at AidKit.
• Evaluation & safety: human-in-the-loop eval harnesses, agentic RAG observability, LLM red teaming (OWASP).
• Technical education: curriculum development and hands-on technical education, building labs with telemetry and custom AI-enabled production learning platforms, grounded in andragogy and heutagogy (backward design, constructive alignment, ATD, Kirkpatrick evaluation).

• What We Will: building a community platform that connects laid-off workers to legal, healthcare, and job-search resources. Piloting an evaluated agentic RAG system for labor-law citations. Repo: layoff-qa-pipeline.

• AI safety research (in progress, first-author publication, 2026): measuring whether post-training quantization and GGUF compression widen the MultiJail high- vs low-resource jailbreak gap (Δ_HL) on open-weight models, under matched decoding and a dual-judge ASR protocol, alongside layer-localized safety interventions at the transformer-layer level. Repo: quant-multilingual-safety.

• AI & labor policy research (in progress, 2026): collaborating with TechEquity on a study of survey data from tech workers about how generative AI is changing their roles, skills, and job quality.

• Newline AI Research Track
  • Implementing decoder-only Transformers in PyTorch: multi-head self-/cross-attention, RMSNorm/LayerNorm residuals, SwiGLU FFNs, RoPE, and KV-cache; pretraining GPT-2 124M with DDP and gradient accumulation
  • PEFT instruction tuning with Hugging Face PEFT: LoRA/DoRA, BitFit, and prompt tuning for domain and format control
  • Embedding and multimodal adaptation: CLIP-style alignment, triplet and contrastive losses, and hard-negative mining
  • Preference and RL post-training: SFT into RLHF/PPO critic-actor loops, DPO, and GRPO / Dr. GRPO / GiGPO with group-relative advantages, KL-to-reference regularization, and Math-Verify-style verifiable rewards
  • Layer-localized safety methods: freeze-all-but-k GRPO with layer contribution C(k); Safety Layers / SPPFT (freeze middle refusal layers under capability fine-tuning); ESI/SET/SPA sparse safety-critical updates (~1% of weights); Arditi-style refusal-direction ablation and addition in the residual stream
  • Systems context for eval harnesses: MoE routing, agent tool-use loops, Graph and multi-hop RAG; retrieval stacks (Hugging Face, DSPy, LangGraph); production reliability (SRE SLOs, model orchestration mixing multi-model cloud and local LLMs for cost optimization)

M.S. Computer Science, UPenn · Post-grad certificate in Machine Learning & AI, MIT · MPA-BA Data Science for Public Policy, NYU (Reynolds Fellow) · A.B. Mathematics-Economics, Columbia (Centennial Scholar)

💬 Reach me at kaitlin.zhang@owasp.org or kaizencode.art.