The situation
In 2016, TVT’s Asia office in Singapore was a small part of a global media services company headquartered in the UK. The office prepared programs for broadcasters in Asia: editing to fit time slots, checking compliance with local broadcasting rules, and adding subtitles and dubbing. For each client, TVT set up a customized version of its workflow system, ContentSelect.
The Asia office was growing and wanted to become more self-sufficient, which included technical expertise. Until then, every client project needed someone from the UK to work out what the client needed and how to connect ContentSelect to the client’s systems. That was slow and costly.
TVT Asia’s director needed a technical counterpart in Singapore. A friend of his from the startup community had seen my work at a workshop he’d organized for a Singapore media company and introduced us. I started in 2016.
What I did
Media projects for major broadcasters
I joined the planning discussions with each broadcaster in pre-sales and defined the technical solution. That included the media workflow from broadcast schedule to delivery, with tight deadlines so every program was ready in time for broadcast. It also meant integrations: schedules and program details came in by spreadsheet or API, and the media itself by large file transfer or on physical drives. I worked out what to reuse from past projects and what to custom build, estimated the cost, and led the implementation with TVT’s developers in Poland.
Most projects followed a similar pattern: editing to fit time slots and formats, compliance checks for each country’s rules, quality control, and delivery back to the client. Often clients had additional custom needs. For one broadcaster, we built a preview platform where their teams and invited journalists could watch upcoming programs before broadcast.
Over the years, I worked on projects for nine major broadcasters and streaming services.
A platform to manage 8,000 linguists
When I joined, TVT tracked about 200 linguist contractors in one large Google Sheet, with their personal and bank details and any problems. Coordinators onboarded each new linguist by email, step by step. How many linguists the coordinators could manage by hand limited how much work TVT could take on.
In 2018, with a few of TVT’s developers, I built a prototype to replace the spreadsheet: Linguist Profiles, a structured database with a simple interface. Later, we moved development to a contractor in Vietnam that I had brought in for TVT’s projects in Asia in 2017. Over the following years, I turned the prototype into a full platform.
The platform took over the routine work. Onboarding became an automated workflow in several steps: personal and payment details, references, signed agreements and guidelines, and one or two assessments. Coordinators stepped in only for exceptions, such as a failed assessment or a special requirement.
Once onboarded, linguists set their availability, saw their tasks, and received client feedback on their work. The system created a TVT email address for each linguist and provided a team chat, both under a single sign-in. With thousands of linguists on the platform, services priced per user were out of the question, so we self-hosted the tools for sign-in, chat, and email.
TVT expanded step by step. The number of linguists grew from about 200 to 8,000, while the coordinator team stayed the same size.
Retiring the server room
When I joined in 2016, TVT ran its own server room: equipment to ingest media from clients’ drives and tapes, powerful machines to convert media, and tape storage for archived media. That had long been necessary, since master files run to dozens of gigabytes, too large for the internet connections and cloud services of the time.
In 2018, TVT decided to move to the cloud (AWS). The system administrators handled the infrastructure. I led the software side: making TVT’s systems work in the cloud and moving workflows off the old servers.
We moved from the least connected parts to the most connected. Archived media went first and media processing last. The administrators had built an impressive system of folders and scripts: a file dropped into a certain folder was processed, renamed, and passed on to the next step. We moved these workflows one at a time, some into ContentSelect and some into cloud functions. Users then started them from ContentSelect instead of moving files by hand and knowing which folder each one belonged in.
The move took about half a year and allowed TVT to stop maintaining old hardware and move into a smaller office. Savings in hardware, maintenance, electricity, office space, and administration came to an estimated $300,000 a year. Nobody was let go: the administrators moved on to running TVT’s systems in the cloud.
AI for subtitles and compliance
Subtitling looks like translation, but broadcasters’ strict style guides for line length, wording, formatting, and more are hard for classic machine translation to follow. Names and terms must also stay consistent across episodes, seasons, and translators. From 2023, I led work on generative AI that drafted subtitles following the broadcaster’s style guide and the show’s glossary. These drafts then served as a starting point for human subtitlers to review and finish.
The AI also provided feedback. Human quality control is a separate step by another person, often a day or two later and in another time zone. Our AI review flagged odd wording and terms that didn’t match the glossary right away, on top of the existing automatic checks.
For compliance, AI did a first pass over each program and flagged sensitive scenes, such as drug use or violence, for human reviewers. In both cases, the final work and the decisions stayed with people.
Outcome
By 2024, TVT’s business had shifted. In 2016, media management for broadcasters was the main business and subtitling a small part. With the coordinators no longer the limit, it became the reverse: the network of subtitlers had grown 40 times to 8,000 linguists, managed by a team of fewer than ten, about the same size as before.
Lessons
What worked
Automating routine work was a key driver of efficiency at TVT. ContentSelect moved files through repeatable steps and kept each project’s information in one place. The cloud move replaced folders and scripts people had to know by heart, and the linguist platform automated onboarding, task management, and invoicing. People focused on the work that needed judgment and on the exceptions the systems flagged.
ContentSelect was built as a reusable core that TVT deployed and customized for each client. Workflows, integrations, and screens were often similar between clients, so we could set up instances for new clients quickly. Once set up, a client’s system often ran for years, changing only when the client requested changes to their workflow or data.
Serving thousands of linguist contractors, we needed sign-in, chat, and email for all of them. None of these was core to TVT’s business, so we didn’t want to build them, and per-user pricing ruled out commercial services. We self-hosted open-source tools instead: Keycloak for sign-in, Zulip for chat, and our own mail server with Horde for webmail. It was the right call: we only had to build the integrations and we avoided per-user fees that would have run to thousands of dollars a month.
What I’d do differently
Each client’s ContentSelect started as a copy of the core or of another client’s version with similar requirements. Customizations passed from one client to the next but rarely made it back into the core. Setting up a new client instance took someone who knew all the versions to find the closest starting point, and sometimes meant piecing together work from versions that had grown apart. Today, I’d keep one shared core and merge new features back into it, or offer them as plugins any new client instance can use.
Running our own sign-in, chat, and email saved subscription fees, but keeping these tools running and up to date came with its own cost and effort. With so many people in the system, security updates were critical, and anything that broke had to be fixed at once. We relied on an outside specialist, so urgent fixes depended on their availability and could cost extra, while the knowledge sat outside TVT. Today, I’d make a greater effort to bring that expertise in-house and reduce vendor dependence in business-critical areas.
For anything ContentSelect or the linguist platform didn’t already do, we worked out the requirements with TVT’s internal teams through mockups, going back and forth. A prototype people can click through would have shortened that a lot, but building one was often not worth the effort. Today, I’d use AI-built prototypes: they’re quick to build and help everyone agree on the requirements early.