AI Developer (LLM Applications & Data Pipelines)

About us Marketsoft is a Sydney-based data management company. We deliver data and marketing technology projects for clients across Australia. We're building a small, vetted pool of freelance specialists we come back to repeatedly — not one-off gig hires. If you do good work with us, you get more of it. What you'll work on We're applying AI to real operational problems in data delivery, not to demos. Typical work includes: 1. Building LLM-driven steps into existing data pipelines — for example, automated data health checks that flag anomalies in client datasets before they reach delivery 2. Designing and hardening prompts so outputs are reliable enough to run unattended in production Integrating model APIs (Anthropic, OpenAI and similar) into internal tooling 3. Handling structured output — getting consistent, parseable JSON out of models and validating it 4. Building evaluation and QA around AI steps so we know when something has drifted What we need from you 1. Demonstrable, hands-on LLM application work — shipped, not experimental 2. Strong Python 3. Prompt engineering and prompt-chaining, with real judgement about where models fail 4. Experience calling model APIs and wiring AI into a larger pipeline rather than a standalone chat interface 5. Structured-output handling (JSON) and validation 6. Evaluation and QA of model outputs — you can tell us how you know an AI step is working 7. Clear written English, and a habit of asking clarifying questions before building the wrong thing Nice to have 1. Experience embedding AI into reporting or data-quality processes 2. Agentic workflows 3. Familiarity with data platforms and marketing technology stacks Important — data handling You'll be working with client personal data across multiple jurisdictions. You'll need to sign an NDA and follow our data-handling rules. To apply — please answer these (We read these properly. Generic proposals will not be shortlisted.) 1. Describe an LLM feature you took to production. What broke, and how did you make it reliable? 2. Here's a badly written prompt: "Look at this data and tell me if anything seems wrong with it." Tell us what's wrong with it and how you'd rewrite it for a production pipeline that must return consistent, parseable results. 3. How do you evaluate whether an AI step is working correctly once it's running unattended? 4. What's your Python experience, and what's the most data-heavy pipeline you've worked on?