ML ENGINEER - VOICE PIPELINE & DATA ARCHITECTURE

• Full-Time
• Remote
• 3.000 USD

PROBLEMS WE'RE SOLVING

1. Data pipeline isn't architected for scale or evolution. We have a working voice model, but no systematic approach to gathering better training data, experimenting with new models, or understanding what's limiting our performance. Data collection is ad hoc. Model iterations are slow.

2. Inference cost and speed are bottlenecks. Our current model works, but we don't know if we're paying more than we need to, if it's slower than it could be, or if smaller, faster models could do the same job. No one is optimizing for inference efficiency.

3. Infrastructure is more complex than it needs to be. The pipeline works, but it's not designed with clarity. Too many moving parts, unclear data flow, operational complexity that costs money and attention.

4. No one is thinking structurally about what comes next. Where do we go from here? What data would unlock better predictions? Can we train on specific cohorts? Should we have multiple specialized models instead of one general one? These questions aren't being asked.

WHAT THIS PERSON DOES

Must have

- Have agency on evolution. You decide what the next model should be, what we should measure, where we should experiment. This is your strategy to own.

- AI-forward thinking: You work with AI tools, as part of your problem-solving toolkit. You use them to prototype, debug, explore data, and accelerate iteration. You understand their strengths and limitations. You're not afraid to experiment with AI assistance for model development, data pipeline work, or architectural thinking. You know when to use them and when to reach for traditional approaches.

- Own the data pipeline architecture end to end. Design how we collect, validate, label, and version training data. Build systems that scale as our data grows. Understand what data gaps are limiting model performance and how to fill them.

- Build and maintain production models. This is your work, not someone else's infrastructure. Train new models, evaluate them against production, decide when to ship. Understand why a model works and what breaks it.

- Think architecturally about the whole system. How does data gathering inform training? How does training inform serving? How does serving cost drive data strategy? You're not optimizing one piece; you're optimizing the system.

Could have

- Optimize for inference. Benchmark our current model on speed and cost. Experiment with quantization, distillation, smaller architectures, anything that gets us better performance per dollar. Own the inference pipeline: how data flows from Deepgram → model → output.

- Reduce complexity and cost. Look at the current pipeline, the infrastructure supporting it, the operational overhead. Simplify. Move away from things we don't need. Make the system smaller and faster, not bigger and slower.

HARD SKILLS REQUIRED

Must have

- ML fundamentals and production experience. You've trained models, evaluated them, shipped them. You understand PyTorch or TensorFlow deeply. You know the difference between a model that works in notebooks and one that works in production.

- Audio and voice processing. You understand audio pipelines, speaker diarization, audio feature extraction, speech-to-text APIs. You can debug when Deepgram output is wrong, when audio preprocessing is the bottleneck, when we're losing signal in the pipeline.

- Data engineering and pipeline design. You can design systems to collect, validate, version, and serve training data at scale. You understand data quality, labeling workflows, and how bad data breaks everything.

- Model optimization. You've done inference optimization: quantization, pruning, distillation, architecture search. You know how to make models smaller and faster without losing accuracy.

- Python (primary). This is where the work lives: training scripts, data pipelines, experimentation. You're fluent here.

- SQL and data fundamentals. You read and write SQL. You can query what's in our databases and understand what data we have to work with.

Could have

- Infrastructure and deployment (secondary). You don't need to be a DevOps engineer, but you need to understand how models move from training to production, how they're monitored, what happens when they fail. You know Modal, containerization, async jobs.

SOFT SKILLS / BEHAVIOR EXPECTATIONS

Must have

- Architectural thinking over parching now, fixing later. You see a problem and you think about the system that created it, not just the immediate fix. When inference is slow, you're not just tweaking the current model; you're asking if the architecture is right.

- Bias toward simplicity and clarity. You prefer a pipeline you can trace versus one that's theoretically optimal but nobody understands. You remove steps, not add them.

- Agency in uncertainty. You make decisions when data is incomplete. You run experiments to answer questions, not ask for permission to start. You're comfortable saying "this is the direction we should go" and explaining why.

- Rigor about measurement. You don't ship a model without understanding how it compares to the current one. You benchmark inference speed, cost, accuracy. You know what you optimized for and what you might have lost.

- Curiosity about the real constraints. You ask: What's actually limiting us? Cost, latency, accuracy, data? You dig into the numbers before deciding what to optimize.