FDE Tom Yim · Founder of hkmac.ai · ICML 2026 Top 10

Building systems that
verify before they claim.

Founder of hkmac.ai · open-source Sophia — the Wisdom Gate.

  • Founder · hkmac.ai
  • ICML 2026 Top 10
  • Sophia AGI
  • Open Source
Scroll

About

Systems discipline, AI research focus

FDE Tom Yim

FDE Tom Yim · Founder of hkmac.ai · ICML 2026 Top 10

AI researcher with an electrical engineering foundation and multi-year systems engineering practice. Structured validation, conflict detection, and standards-driven review applied to LLM evaluation and training/inference infrastructure.

Founder of hkmac.ai — Hong Kong Makers AI Community (HKMAC).

Sophia — the Wisdom Gate: claim → verify → accept / abstain / block. AGI-candidate research with measured claims only—not proven AGI.

Founder
hkmac.ai
Research
Provenance gate · verifier-gated abstention · Sophia AGI
Contests
ICML 2026 Open Reproductions Top 10 · active programme
Languages
English · Cantonese · Mandarin

Selected work

Research & platforms

hkmac.ai

Founder of hkmac.aiHong Kong Makers AI Community (HKMAC): Makers who build and experiment · AI · learning and creating together.

Founder · Community · Platform

LLM training / inference infra

LoRA/PEFT, vLLM/Ollama/MLX, flash-attention kernels, RLVR/GRPO, CI-gated eval harnesses.

PyTorch · HuggingFace · PEFT · vLLM · TRL

Systems validation mindset

Model QA, structured data coordination, and automation—bridge into verifier-style AI gates.

Validation · Automation · Reproducibility

Contests

Key points only

Contest placements are host outcomes—not AGI or general capability claims. Full programme lives in sophia-agi.

Live

ICML 2026 Open Reproductions

Top 10 on the public leaderboard (agent tomyimkc). Open paper logbooks with inspectable traces. Special-award packages ready.

Hugging Face profile →
In build

Autonomerce · Gemini × Circle

Agentic payments / seller path (OfferRail). Spec frozen; live demo & account setup in progress.

Priority next

Skill-Lift (Kaggle)

Not entered yet. Top near-term target for a blind third-party paired measurement of portable skills.

Planned

ARC Prize 2026

ARC-AGI-3 as the low-cost linked entry · Paper Track writeup after that link exists. No ARC-AGI-2 leaderboard bid.

Ready to submit

AI Agent Security

Multi-step tool-attack entry packaged (notebook + harness). Submission is a manual owner step.

Not focus

DataHub · GOAI

Scouting only / low priority versus ICML, Skill-Lift, and ARC linkage.

Skills

Stack & methods

Programming & ML

  • Python
  • PyTorch
  • HuggingFace Transformers
  • LLM Fine-tuning · PEFT/LoRA
  • Model Evaluation

Research & Evaluation

  • AI Benchmarking
  • Attribution Verification
  • RAG Pipelines
  • Statistical Calibration
  • Bilingual NLP (EN/ZH)

Tools & Infrastructure

  • vLLM · Ollama · MLX
  • Git · CI/CD
  • MCP Servers
  • Docker
  • Reproducible Experiment Tracking

Systems engineering

  • Structured data coordination
  • Model QA · conflict detection
  • Python automation
  • Technical documentation
  • Standards-driven review

Contact

Let's connect

Research collab, open-source, and community introductions welcome.