We are looking for a Senior Machine Learning Engineer to own our voice ML system end to end in a fast-moving, low-bureaucracy environment.
We have a working voice model in production, and its outputs feed a matching layer between candidates, roles, and teams. The next step is making the whole system more measurable, scalable, and easier to evolve. You will own the full path from training data to model evaluation to inference, and help shape how matching learns from real hiring outcomes.
This is not a research-only or MLOps-only role. You will work across data pipelines, audio preprocessing, model training, evaluation, serving, and production debugging.
The role is a strong fit for someone who can take a working system, find what is actually limiting it, run the experiment that answers the question, ship the improvement, and prove it with numbers.
What you will do
- Contribute to the matching layer that turns behavioral signal into candidate-to-role and candidate-to-team fit, making it explainable, testable, and calibratable as outcome data grows.
- Own the training data pipeline: collection, validation, labeling, versioning, and dataset creation.
- Train, evaluate, and ship models, benchmarking every candidate against the current production model before deciding what goes live.
- Define what gets measured across accuracy, latency, reliability, and failure modes.
- Debug the full voice pipeline, from speech-to-text output through audio preprocessing, diarization, feature extraction, model output, and downstream predictions.
- Optimize inference speed and reliability, including quantization, distillation, or smaller architectures where they improve the system.
- Simplify the ML infrastructure so the pipeline is easier to trace, debug, and evolve.
- Work closely with our psychometrician to translate measurement requirements into model targets, validation criteria, and evaluation design.
What we are looking for
- Experience building predictive or recommendation models on top of extracted features, ideally where labels are scarce, delayed, or noisy.
- Strong production ML experience: you have trained, deployed, and maintained models that real users depend on.
- Deep experience with modern ML frameworks (PyTorch, TensorFlow, JAX) and strong Python for training, data pipelines, and experimentation.
- Experience designing ML data pipelines and a sharp sense of how bad data limits model performance.
- Hands-on experience with audio, speech, or voice systems, including speech-to-text APIs, diarization, and audio feature extraction.
- Experience with model optimization and inference efficiency, and the ability to benchmark accuracy, latency, and performance tradeoffs.
- SQL skills to inspect what data exists and what is usable for training and evaluation.
- Ability to think architecturally across data, training, evaluation, serving, and operations.
- Clear communication about tradeoffs, risks, assumptions, and progress.
Nice to have
- Experience with speech-to-text (STT) providers and APIs.
- Cohort-specific training, specialized models, or model routing.
This role is probably not a fit if
- You need highly detailed specs before starting.
- You prefer slow-moving enterprise structure.
- You only want to train models and hand off serving, or only want to run infra and never touch the model.
- You ship models without comparing them to production.
- You are uncomfortable making decisions with incomplete data.
- Use AI-generated code without verifying it.
Within 90 days, success means the engineer has shipped at least one measurable improvement to the voice model, its pipeline, or the predictions it feeds, from problem definition to production. They should be able to show the before/after evidence, explain the modeling and data tradeoffs, and leave the pipeline easier to operate or debug. They should have worked across data, training, and inference, handled at least one ambiguous request with limited guidance, and demonstrated safe, practical use of AI tools in their workflow.
What we need you to tackle
We are looking for a high-ownership ML engineer who can make a production voice model more reliable, measurable, and easier to evolve with limited hand-holding. Strong candidates can explain what they personally built, what tradeoffs they made, what broke, and how they measured whether the model actually improved. They should be comfortable across training data pipelines, audio and speech processing, model training and evaluation, inference, and production debugging. They should use AI aggressively as leverage for prototyping, data exploration, and debugging, but not blindly trust its output. Weak fit: someone who needs clean labeled data handed over, optimizes metrics without questioning the label, works only in research settings, or needs fully defined specs to make progress.
Data Modeling
Data Science
Machine Learning
Python
Data Engineering
Day to day
- Own the training data pipeline end to end: collection, validation, labeling, versioning, and dataset creation.
- Debug the full voice pipeline, from speech-to-text output through audio preprocessing, diarization, feature extraction, and final model output.
- Run experiments comparing new models against production and decide what ships based on measured accuracy, latency, and failure modes.
- Translate ambiguous questions from founders, product, or Ops into small executable experiments.
- Simplify the ML infrastructure so training and inference are easier to trace, debug, and maintain.
Nice to haves
- Optimize inference speed and reliability via quantization, distillation, or smaller architectures while managing accuracy-performance tradeoffs to meet SLOs.
- Trace errors end-to-end—from STT output through preprocessing, features, model predictions, to downstream matching to isolate root causes quickly and prevent recurrence.
- Deep skill with modern ML frameworks to build and iterate efficiently on models powering voice and matching (e.g., PyTorch/TensorFlow/JAX).
- Design predictive/recommendation models robust to scarce, delayed, or noisy labels using appropriate objectives, sampling, or weak supervision strategies.
- Use SQL to discover data availability, cohort differences, label delays, and issues that affect training and evaluation readiness.
- Design models and training strategies that tolerate scarce, delayed, or noisy labels (e.g., weak supervision, robust loss, confidence learning).
- Structure data, training, evaluation, serving, and operations into a traceable, evolvable pipeline that reduces complexity and speeds iteration.
Company
We thrive in the technology industry by using AI and voice analysis to connect diverse talent with opportunities. Our innovative approach saves time and costs, fostering a dynamic work environment where unique perspectives are celebrated and growth flourishes.
Senior Machine Learning Engineer, Voice AI Brasil