Anthony @antmikinka

About

Anthony Mikinka

Data science undergrad and ex-AMD co-op building local-first AI systems: multi-modal LLM apps (Lemonade Interviewer), multi-turn agent reliability, domain prompting under constraints, and multi-agent orchestration (LOOM). Strong in TypeScript/React, Python, agent architectures, and hardware-aware optimization for Ryzen AI and edge deployment.

I'm Anthony Mikinka (@antmikinka) — a software engineer focused on local AI, open source, and tools that make on-device models practical.

Co-creating a reality that is aligned with my higher selves, ambitions, and the most amount of freedom

I previously worked at AMD (ex-AMD) as an undergrad co-op on agent task planning for resource-constrained client devices—including Lemonade Interviewer and multi-turn reliability research under Ryzen AI / client hardware constraints. The same problems still drive my work: local-first agents, context budgets, and systems that stay honest on real devices.

Education: B.S. Data Science, Arizona State University (Expected 2027) · based in San Jose, CA.

Education

Experience

  1. Jan 2026 · ex-AMD

    Undergrad co-op — Agent task planning for resource-constrained client devices

    Advanced Micro Devices (AMD)

    Designed and demonstrated Lemonade Interviewer for Lemonade SDK: open-source local-first multi-modal interview agent (TypeScript/React) on Lemonade Server with Whisper ASR, local LLM reasoning, and on-device speech-out. Researched multi-turn reliability, domain prompting under constraints, and validation on Ryzen AI / client hardware.

    • Lemonade Interviewer: dynamic personas, multi-phase sessions, graded feedback under client constraints
    • Showed multi-turn tool selection and goal completion degrade without compression, checkpointing, or history pruning
    • Domain-specific instructions recovered capability on hard multi-step tasks (e.g. Newton’s Cradle sim) for small models
  2. Mar 2022 – Nov 2023

    Data Processing Coordinator

    ArborOakland Group

    Constraint-aware data pipelines and automation under hard budget ceilings, compliance requirements, and delivery SLAs.

    • Cut mailing costs ~15–30% while maintaining compliance and delivery SLAs
    • Automated financial reporting and job costing for 200+ client projects (~70% less manual effort)
  3. Jun 2021 –

    Founder

    Symbiosis Hemp Clothing

    Founded and operated a sustainable product business end-to-end; applied automation and agent-oriented orchestration ideas to production process management.

    • End-to-end product, manufacturing, and supply-chain operations
    • Automation workflows under real operational constraints
  4. Ongoing

    Independent builder & open source

    Personal / community

    Research (UCL), multi-agent systems (LOOM), edge/NPU kits, Hugging Face edge models, and contributions such as BMAD-METHOD agent templates.

    • Universal Conditional Logic (arXiv:2601.00880) — 29.8% token reduction (p<0.001)
    • Open-source contributor to BMAD-METHOD (agent templates; project later grew past 51k stars)
    • Published 11+ optimized models on Hugging Face for edge deployment

What I work on

On this site

Want to work together?

See Contact for the best channels, or browse projects and what I’m doing now.