Technical Operations Intern — AI Inference Platform

Who we are

TMax is an AI inference platform built for Asia. We buy model capacity straight from Chinese labs — ByteDance/Volcengine, Alibaba, Kuaishou — at origin pricing, and serve it through one OpenAI-compatible API. 50+ models across video, image, text and speech, with video as our core market. On top of that we ship white-label apps for companies that want a finished product rather than raw tokens.

Small team, real customers, a lot of ground to cover.

---

What this actually is

In a given fortnight you might price a new video model, dig through a supplier contract for the clause that changes what we're allowed to resell, rewrite a section of the website, help a customer get their integration working, chase down why a pod won't start, and put together a deck for Friday.

You won't know how to do most of that on day one. What matters is that you figure it out and come back with something finished.

---

What you'd work on

Platform — our Go API gateway and the adaptors that connect each upstream provider, onboarding new models and getting their billing right, the Kubernetes cluster everything runs on, test sweeps across the image and video models we serve, and shipping features on top of the API.

Commercial — pricing and margin models across the catalogue, competitive benchmarking against comeptitors, supplier contract review, billing and payments reconciliation across Stripe and Airwallex, and the website, decks and onboarding docs that go out to customers and partners.

The two sides feed each other, which is the whole reason this is one role. The person who onboards the model is the person who knows what it costs, and the person who reads the supplier contract is the person who knows what we're allowed to build.

---

What we're looking for

Three things we won't compromise on:

You work independently. We'll tell you what needs solving, not how. Decide anything reversible yourself and only bring us the things that genuinely need a call. When you bring something, lead with the recommendation and put the reasoning underneath — we should be able to act on it in a few minutes, not read twenty pages and guess what you think.

You'll go anywhere. Engineering one day, contracts the next, a landing page after that. If some part of that makes you think "not my thing," this isn't the role.

fYou use agent harnesses properly. Claude Code, Codex, OpenCode — whichever, but daily and fluently, including the config layer (instruction files, custom commands, skills, MCP, hooks). More importantly: you know what to check before it ships. Agents are confidently wrong all the time, and your work touches billing and production.

Also required:

• Mandarin, written and spoken. Supplier docs, consoles, contracts and support are largely Chinese-only, and our collateral goes out bilingually.
• Good written English — the decks and briefs go to enterprise buyers.
• Proficient Programming in any Language. Any language. You need to read a diff and tell whether the approach makes sense, not just whether it runs. You do not need to be a strong engineer — the bar is "can direct and check," not "can build unaided."
• Comfortable in a terminal. Git, logs, env vars, running something locally.
• Solid with spreadsheets. You'll be expected to deliver documents to help facilitate decision-making in a short period of time with less help from agents when dealing with spreadsheets.
• You learn fast. Go, Kubernetes, Docker, contract review — pick them up as you go. Prior experience is a bonus, not an expectation.

Nice if you have it: exposure to APIs and usage-based billing, some sense of the Chinese model ecosystem (Doubao, Qwen/Wan, Kling, MiniMax, DeepSeek), an eye for design.

---

What you'll own

Real ownership, not shadowing. Two halves, both yours to run:

Technical

The gateway and its adaptors. Our gateway (Go) normalises requests across every upstream provider and bills for them. Providers handle multimodal input very differently — some fold any number of images into a content array, some take only the first and quietly drop the rest, some have no images field at all. Knowing those differences across the Doubao, Vidu, Alibaba, Jimeng, Kling, Hailuo, MuleRouter and Gemini/Vertex adaptors is a big part of the job.

Model onboarding, end to end. Read the provider's docs (often Chinese-only), test the endpoint, map the request and response shapes, extend the adaptor, register the channel, configure the billing, and verify against real call logs before it goes live with assistance from the tech team.

Billing configuration. Per-token ratio, tiered expressions, per-call fixed price, upstream-response passthrough — including resolution-tiered video billing. When the billing system can't express what commercial needs, you get it built.

Infrastructure. The Kubernetes estate (k3s under Rancher): Deployments, Services and targetPort routing, ingress, namespaces, secrets, and the reverse-proxy gateway in front of it. Building and publishing Docker images. Debugging production to root cause when a 502 shows up.

Model QA. Test sweeps across the full catalogue on payload validation, field normalisation, timeouts and output quality. Then reports that separate our bugs from upstream behaviour we need to document and price around.

Business

Pricing and margin. The cost-to-price model across the catalogue: landed supplier cost, markup, customer price, margin per call. Plus competitive benchmarking against competitors, normalised properly — /image, $/1M tokens split in and out, sources cited. Flag it early when margin starts eroding.

Supplier contracts. Read them and pull out what actually constrains us — resale restrictions, minimum commitments, platform-operator obligations, price-adjustment rights — and make sure our customer-facing terms line up.

Billing ops and reconciliation. Prepaid wallet and PAYG mechanics, top-ups across Stripe and Airwallex, and tying payment rails to wallet credits to gateway logs to the books when something doesn't match.

The website and everything that goes out. Model catalogue, pricing presentation, copy. Plus decks, partner briefs, API onboarding guides and one-pagers — in English and Chinese, to a standard that goes to an enterprise buyer without anyone rewriting it.

Customer work. Helping trial accounts get integrated, and turning what they tell you into pricing and product signal rather than testimonials.

---

Why bother

You'll touch the whole business — pricing, contracts, infrastructure, product, customers — in a market that's moving faster than almost anything else right now. You'll be trusted to make calls rather than asked to draft options. And you'll come out of it able to point at things that actually shipped.

Strong performers have a route to a full-time offer.