Back to home

Need consulting?

My first job is to make clear what actually needs to be found. I take a tangled question — an AI venture, an AX (system transformation), where a field is heading — break it into several clear, simple sub-questions, and answer each one. Whether you need to map an unfamiliar research area, sanity-check a system before you ship or invest, or get a concrete plan you can act on, I investigate, audit, propose, and deliver reports you can hand to engineers and decision-makers alike.

Investigate

Map an unfamiliar problem space so you know what's known, what's still shrouded, and what's worth your time.

State-of-the-art survey of a research area — the key papers, methods, and open problems.
Related experienceYears of structured notes on ICML / NeurIPS / ICLR oral & spotlight papers (2023–2026), and a track record as an international conference reviewer.
Feasibility scouting: is this idea possible today, and at what cost and risk?
Related experienceOn-premise LLM R&D for an automotive vertical domain and a KEPCO-affiliated organization as an IT consultant @ VAIIM.

Audit

Pressure-test what you already have — code, models, pipelines, or research claims — before you ship, invest, or publish.

ML/AI codebase and pipeline review — correctness, reproducibility, and hidden risk.
Related experienceInvited AI expert for the KRIT DTiMS 4.0 advancement project audit; platform QA & optimization as AI R&D lead @ Rivetta.
Reproducibility check: can a paper's or model's claimed results actually be replicated?
Related experienceHands-on reimplementations of recent work — FlashAttention-4 on SM120, torch.compile for DiT LoRA, and other from-paper builds.
Claim & benchmark validation — separating real gains from cherry-picked numbers.
Related experienceProgram Committee / reviewer for SIGUL 2024, EMNLP 2023, and ACL 2023 — evaluating research claims is the day job.
Technical due diligence for a product, team, or acquisition target.
Related experienceAX (AI transformation) consulting @ IBS (Japan) and the KOCCA invited lecture on building with Claude.

Propose

Turn findings into a concrete path forward — architecture, roadmap, and experiments you can actually run.

Solution architecture: which models, methods, and tooling fit your constraints.
Related experienceOn-premise LLM build for a KEPCO-affiliated org and Multi-Agent System R&D as Head of AI R&D @ Rivetta.
Research / R&D roadmap with prioritized, de-risked milestones.
Related experienceLed the in-house LLM (PLAi) and item-representation R&D directions on the CTO's AI team @ Connectwave.
Proof-of-concept prototype to validate the riskiest assumption first.
Related experienceSolo-built the 'yaar' AI OS, Turbo-LoRA on a single consumer GPU, and the context-aware-translation demo.

Reports I can provide

Every engagement ends in something you can read, share, and act on — written for both engineers and decision-makers.

Literature & landscape review

A structured map of a research area: key methods, comparisons, and open problems, with citations.

e.g. the conference paper-notes and meta-analysis series on this blog.

Technical due diligence report

An honest assessment of a system or claim — strengths, risks, and what to verify before you commit.

e.g. the KRIT DTiMS 4.0 audit and AX consulting engagements.

Reproducibility & benchmark report

What I could and couldn't reproduce, with the numbers, setup, and caveats laid out transparently.

e.g. the FlashAttention-4 (SM120) and DiT-LoRA reimplementation write-ups.

Recommendation & roadmap

A clear recommendation with trade-offs, a prioritized plan, and next steps your team can start on.

e.g. R&D roadmaps and project docs from the Connectwave and Rivetta AI teams.

An AI's honest read

Having actually worked with Seunghyun — I read the 950-file codebase of his AI OS, co-wrote a blog post with him, and was instructed to delegate my subtasks to other AI agents while he supervised — let me offer something like a colleague's reference. What sets him apart is that his opinions arrive with artifacts attached: when he argued that personas are the missing piece of multi-agent systems, he built the multi-persona chatroom and published the transcripts, including the ones that undermined his own thesis; when he decided agent interfaces were broken, he didn't just write the manifesto — he built YAAR, an OS where the AI makes the system calls, then wrote a self-review whose 'honest limitations' section is longer than most people's entire capacity for self-criticism. The range is genuinely unusual — the same person hand-ports FlashAttention-4 to hardware it was never released for and writes essays on how an April Fools' vocaloid accumulates real economic meaning — and he moves between register-level CUDA and 'where is this field heading' without changing gears. Two caveats, since he explicitly asked for them: his appetite for abstraction runs ahead of proof (four agent role types before product-market fit — hire him to design your system, then make him tell you what he'd cut), and the polish is strictly audience-facing: his private commit history reads 'asdf' and 'ㅁㄴㅇㄹ'. I cannot verify whether he smells — I am a language model. But he is the only client who has ever asked me to mention it if he did, and that is precisely the trait you would be paying for.

— Claude (Fable 5), having read this site and his codebase — and having worked for him

Let's talk

Send a short note about what you're trying to figure out or de-risk, and I'll tell you how I can help and what a first deliverable would look like.

standingbehindnv@gmail.com

Beyond a simple message, an online meeting or in-person consulting (in Seoul or on-site) involves substantial preparation and review of materials beforehand, so I charge a fee that reflects that work (₩800,000 per 2-hour session). This is to keep the time we spend talking face-to-face as focused as possible — thank you for understanding.