I deploy private AI systems that never send your data to the cloud. The whole thing runs on hardware you control: your premises, your server, your rules. I work remote-first, worldwide. I spec it, build it (by directing LLMs, verified at every step) on your infrastructure over secure remote access, and hand you a system plus the runbook to operate it, a documented handover your own team can run.
Email me about a project LinkedIn
Built for law firms, clinics, financial and government-adjacent teams bound by data-protection rules or internal compliance. Deployments are PDPA/GDPR-aware, which matters to any organisation handling Singapore or EU personal data. Also for studios and agencies who want local video, voice, and avatar pipelines without per-seat SaaS bills. Currently serving international clients.
A private LLM running inside your facility. Your data physically cannot reach a cloud API: enforced by routing, not by a promise.
vLLM or llama.cpp on your hardware (or I spec the hardware to buy), with fallback chains that fail loudly rather than leak silently, Tailscale-only access, default-deny firewall, monitored backups, and a runbook your team can operate. The exact discipline I run my own stack on: see it, publicly, daily.
Two hours + a written report: what AI can safely do for you, on what hardware, with what data risks.
Gap analysis of your data sensitivity, existing workflows, and cloud exposure → a one-page recommendation and deployment roadmap. Low-commitment way to see how I work; converts to a full deployment if it makes sense.
Object removal, 4K finishing, avatar videos, voice cloning, auto-dubbing. All on-prem.
Turnkey versions of my creative and audio stacks: a webtoon studio's comic-to-video pipeline, a marketing team's avatar ad-reads, a dubbing house's localization. I'm honest about the quality boundaries. The benchmarks are public and prices are tiered by deliverable quality, not hype:
| Tier | What you get | Object removal | Avatar video |
|---|---|---|---|
| Draft | fast preview, source resolution, no finishing | SGD 12/sec (min 80) | SGD 80 |
| Production | QA'd removal + 1080p finish / social-grade avatar | SGD 25/sec (min 150) | SGD 150 |
| Delivery | 4K finish, multi-pass QA, revision included | SGD 40/sec (min 240) | quote |
One product photo in; a spec-checked Amazon/Shopify 3D model out, delivered with a machine-generated compliance report that proves it meets the marketplace's published rules.
Local image-to-3D generation, an automated mesh-ops chain (retopology, UV unwrap, PBR bake, real-world rescale, KTX2 GLB + USDZ), and a compliance gate that refuses assets failing the spec, so a broken model never reaches your listing. The engineering, the honest benchmarks (failures included), and the open-source licence chain are public in the 3D Stack.
I keep the system you (or someone else) deployed alive and current.
Monthly: backup verification, watchdog review, dependency updates, GPU-memory tuning, quarterly failover test. A fraction of a full-time hire, and I document everything, including what breaks.
I'm not a traditional software engineer and I won't pretend to be. I read code fluently and direct LLMs to write it, then verify with gates (the full method). What I bring that most "AI developers" don't: a decade of quality control and security governance (subtitle QC across 36 languages; GRC/SOP authoring at Merck), which is exactly the discipline on-prem AI needs. Validation, guardrails, and the operational habits (monitored backups, default-deny networking, public failure logs) that keep a system running after launch day.
Remote-first and worldwide: I deliver over secure remote access to your infrastructure, work across time zones, and finish with a documented handover. On-site engagements, up to 2-week deployment sprints, are available for overseas clients where travel makes sense.
Cloud AI tools refuse large categories of legitimate professional work: medical and forensic imagery, security research, violence or nudity in a film or fine-art context, legal evidence. Because everything I run is local, there's no content-policy gate and nothing is logged by a third party. If your work keeps getting blocked by a cloud provider's filters, that's a solvable problem. Handled case-by-case, under NDA.