AI Researcher, Voice & Self-Improving Agents

Freya Inc.

Fatih

Sur place

TRY 700 000 - 1 400 000

Plein temps

Il y a 16 heures
Soyez parmi les premiers à postuler

Obtenez une réponse de cet employeur

Votre CV et votre lettre de motivation, adaptés exactement à ce que l’employeur recherche.

Passez les filtres ATS

Résumé du poste

Freya Inc. in Istanbul is seeking an AI Researcher to own research problems across speech models, agent behavior, evaluation, and learning from feedback.

You will train models and collaborate with engineers to deploy results on real calls, with a strong emphasis on learning from failures and improving reliability in a banking context. The role covers speech recognition, speech generation, streaming speech, and data-quality pipelines, with a focus on latency, tool use, and robust evaluation in a

Qualifications

  • Trained or substantially adapted a model or built and evaluated an agent system with clear evidence of contribution.
  • Strong in Python and PyTorch; ability to inspect training loops, tensor shapes, and memory usage.
  • Depth in speech modeling, language models, reinforcement learning, or agent evaluation.

Responsabilités

  • Learn from failures by turning failed conversations and reviewer feedback into training data, and separate model errors from context or tool issues.
  • Build evaluation and simulation tasks to measure whether an agent completes customer requests, including multi-step actions and recovery from tool failures.
  • Investigate how agents choose tools, maintain state, check results, and recover when plans fail; compare approaches on task success, latency, and compute cost.
  • Explore feedback-driven improvement via prompt optimization, supervised fine-tuning, and reinforcement learning where evidence supports them; establish baselines and attribute improvements.
  • Study memory and context: what to retain, what to retrieve, and how to protect customer data during deployment.
  • Build controlled improvement loops from failure discovery to offline evaluation, review, and monitored release.

Connaissances

Python
PyTorch
Speech modeling
Reinforcement learning
Agent evaluation
Experiment reproducibility
Explain research results

Outils

Python
PyTorch

Description du poste

AI Researcher, Voice & Self-Improving Agents
Build agents that learn to do more

Freya started with a goal: help banks serve every customer better. Our long-term vision is an AI Personal Banker that can understand a person's needs and help them act on them. Voice is how people should be able to explain those needs naturally.

A very small team built and released the world's most realistic voice model. We are extending that work across voice, reasoning, tool use, and agents that improve from structured feedback. The research goal is an agent that understands the customer, takes the right action, and becomes more reliable as we learn from its failures.

Our systems serve banks, including deployments inside their own infrastructure. That gives this role a concrete research setting: telephone audio, financial vocabulary, interruptions, variable hardware, and people who expect a useful answer without an awkward delay.

Your role

You will own a research problem through experimentation, model development, and deployment. You will decide how to measure it, inspect the data, train the model, and work with our engineers to make the result useful on a real call.

The scope includes speech models, agent behavior, evaluation, and learning from feedback. Bring depth in one of these areas and a strong interest in the others. You should be able to explain what your experiment demonstrated, where the system still fails, and whether the improvement survives outside your test examples.

This is a full-time, on-site role in Istanbul, working directly with the team building the agents and the infrastructure that serve your models.

Agents that improve through evidence
  • Learning from failures. Develop methods to turn failed conversations, incorrect tool calls, and reviewer feedback into useful training examples and targeted tests. Separate errors in the model from missing context, faulty tools, and ambiguous instructions.
  • Evaluation and simulation. Build tasks that measure whether an agent completed the customer's request correctly. Test multi-step actions, changing instructions, recovery from tool failures, and cases where the right response is to ask for clarification.
  • Reasoning and tool use. Investigate how agents select tools, maintain state, check intermediate results, and recover when a plan fails. Compare approaches on task success, latency, and compute cost.
  • Feedback-driven improvement. Experiment with prompt optimization, supervised fine-tuning, preference learning, and reinforcement learning where the available evidence supports them. Define a baseline and demonstrate which intervention caused the improvement.
  • Memory and context. Study what an agent needs to retain during a task, what it should retrieve, and how stale or irrelevant information affects decisions. Keep customer data and permissions scoped to the relevant deployment.
  • Controlled improvement loops. Build a repeatable path from failure discovery to a proposed change, offline evaluation, review, and a monitored release. Measure regressions as carefully as gains.
Voice research
  • Speech recognition for banking conversations. Improve recognition of names, amounts, dates, and domain-specific terms under telephone compression, background noise, accents, and code switching. Study which errors change the agent's next action.
  • Speech generation that people can follow. Work on pronunciation, pacing, question intonation, and voice consistency. A payment amount or an account instruction must remain clear through a telephone connection.
  • Streaming speech. Investigate how to start speaking sooner while preserving the meaning and delivery of the complete sentence. Measure the effect of chunking, buffering, and interruptions on the experience.
  • Audio representations. Explore codec and speech-language modeling approaches against a defined task. Compare the quality, latency, and compute cost they introduce before committing to a larger training run.
  • Data quality. Build reproducible preparation and evaluation pipelines. Inspect transcripts, alignment, speaker balance, duplicate recordings, and train/test contamination so a reported improvement reflects a real improvement.
  • Deployment performance. Work with our infrastructure engineers to profile GPU memory, inference latency, and concurrency. Measure how changes such as quantization affect the model's behavior as well as its speed.

An agent completes individual tool calls correctly but fails a longer banking task. Design a benchmark that identifies where its state or reasoning breaks, then compare an improved approach against the same held-out cases.

A new prompt fixes the latest complaint but breaks a previously working flow. Build an improvement loop that discovers this before release and measures whether the fix generalizes.

An automated evaluator rewards confident answers even when the underlying action failed. Design checks against the actual task outcome and calibrate the evaluator against human judgments.

A recognition model scores well overall but confuses two monetary amounts. Design an evaluation that exposes the failure, investigate whether data or modeling is responsible, and measure whether your fix generalizes.

A voice sounds excellent in an isolated sample but becomes inconsistent across a longer conversation. Build a listening evaluation that captures the drift and identify the conditions that cause it.

A model meets its latency target with one active caller, then falls behind as concurrent calls increase. Work through the serving and modeling tradeoffs until you understand where the time and memory go.

These are examples of the questions the role exists to answer. You will help choose the highest-value experiments and turn the results into models we can ship.

What makes you a strong fit
  • You have trained or substantially adapted a model, or built and rigorously evaluated an agent system. You can walk through the data, code, experiments, and results yourself.
  • You are strong in Python and a deep-learning framework such as PyTorch. You can inspect a training loop, diagnose unstable optimization, and reason about tensor shapes, numerical precision, and memory use.
  • You have depth in speech modeling, language models, reinforcement learning, or agent evaluation. You can connect a technical decision to a measurable change in what the system does.
  • You make experiments reproducible. You keep useful baselines, change variables deliberately, and investigate regressions instead of selecting only the examples that improved.
  • You can explain a paper's contribution, identify the assumption that matters for our setting, and implement an experiment that tests it.
  • You care about the entire result: whether it is understandable, accurate, fast enough, and feasible to serve in the customer's environment.
Useful experience

Experience with agent benchmarks, tool-using models, preference optimization, reinforcement learning, distributed training, streaming audio, Turkish speech, or inference optimization is valuable. A strong research project with clear evidence of your contribution is useful whether it came from a university, a company, or your own work.

How we work

Research and deployment sit close together at Freya. You will work with the engineers who see failures in customer environments and help turn those failures into specific research questions. You will also help them understand what your model can and cannot reliably do.

The team is small, the quality bar is high, and the pace is intense. We expect persistence through failed experiments, attention to details that affect real users, and the ability to make progress without a complete set of instructions.

Show us a model, agent, evaluation framework, experiment, paper implementation, or repository you are proud of. Explain what you changed, how you measured it, and what you would investigate next.

Obtenez votre examen gratuit et confidentiel de votre CV.

ou faites glisser et déposez votre fichier ici.

Similar jobs

Postes similaires à comparer

AI Researcher, Voice Models
AI Researcher, Voice Models

present of AI • Fatih

Sur place
TRY 3 237 000 - 4 731 000
Forward Deployed Engineer
Forward Deployed Engineer

Freya Inc. • Fatih

Sur place
TRY 300 000 - 540 000
Voice AI Researcher: Self-Improving Agents
Voice AI Researcher: Self-Improving Agents

Freya Inc. • Fatih

Sur place
TRY 700 000 - 1 400 000
EVERYTHING Engineer
EVERYTHING Engineer

Freya Inc. • Fatih

Sur place
TRY 420 000 - 700 000
Forward Deployed Engineer (2 openings)
Forward Deployed Engineer (2 openings)

present of AI • Fatih

Sur place
TRY 600 000 - 1 000 000
DevOps Engineer
DevOps Engineer

Freya Inc. • Fatih

Sur place
TRY 700 000 - 1 100 000
AI Engineer
AI Engineer

Digitopia • Fatih

Hybride
TRY 350 000 - 550 000
Hybrid working
Stock options
Generous paid time off
Voice AI Researcher: Real-Time Speech & Synthesis
Voice AI Researcher: Real-Time Speech & Synthesis

present of AI • Fatih

Sur place
TRY 3 237 000 - 4 731 000
AI Agent Architect & Engineer
AI Agent Architect & Engineer

Wonderful Ltd. • Turquie

Sur place
USD 90 000 - 120 000
Agent Engineer, Turkey
Agent Engineer, Turkey

Wonderful Ltd. • Turquie

Sur place
USD 90 000 - 120 000