About the role

This role owns both halves of applied AI at ContactVerse: choosing and evaluating the models, and building the features that use them. We merged two director roles into one deliberately. At our scale, splitting model selection from model application produces two people arguing about a problem one person could solve. You would own routing, cost, evaluation and agent behaviour, across a provider stack that differs by region.

What you will do

• Own model and provider evaluation. Per region and per capability, against real workloads rather than benchmarks. India runs a different speech and language stack from Europe.
• Own the routing policy. Which class of work goes to which tier of model, what is cached, and what moves to a cheaper model without a quality complaint.
• Own cost per unit of work. Measured, tracked and reduced. Frontier models are the easy default and rarely the right one.
• Own the evaluation harness. Including regression testing of model behaviour, so a prompt change that helps one case does not silently break another.
• Own agent behaviour. Prompts, tools and the context corpus behind them, held under version control, because a correction belongs in the context, not in the single output it produced.
• Own disclosure in the product. Where a reply is AI-drafted, the disclosure evaluation runs before send. That is a requirement, not a preference.
• Ship features, not demos. The measure is what customers notice and keep using.
• Work through Core. No industry vertical calls a vendor directly; capability is built once and consumed by many.
• Keep up without chasing. The provider landscape moves monthly; the discipline is knowing which changes are worth acting on.

Your first year

First 90 days. A model and provider evaluation per region, and a routing policy with real numbers behind it.

By six months. Evaluation and regression testing of model behaviour in place. Cost per unit of work falling.

By twelve months. AI features that measurably work, and the evidence to prove it.

What you will bring

• Eight years or more in engineering, including applied AI or machine learning in production.
• Has shipped LLM features to real users and iterated after the complaints arrived.
• Has benchmarked and selected models against real workloads, and defended the choice with numbers.
• Has moved production traffic to a cheaper model without a quality regression.
• Strong on evaluation: offline, online, and the difference between them.
• Deep practical knowledge of prompting, retrieval, tool use and context management.
• Python, and comfortable in a service codebase rather than only in notebooks.
• Available for a working overlap into the evening IST.

Nice to have

• Speech models (ASR and TTS) as well as text.
• Experience of Indian language or speech stacks such as Sarvam or Krutrim.
• Fine-tuning and when not to bother.
• Multi-provider cost engineering, including buying direct versus through a cloud platform.
• Contact-centre or conversational products.

What you will work with

Python and FastAPI on AWS, a Core-mediated AI layer with a region-scoped provider registry, vendor credentials in AWS Secrets Manager, an evaluation harness under development, and a context corpus held under version control.

What this role is not

Not a research role and not a prompt library. The measure is what customers notice.

How we work

The team works to a UK-aligned pattern, with the main working overlap in the evening IST. These hours are indicative and may change as the business changes; any change would be discussed rather than imposed. Development happens on company-provided remote machines. Everything is remote and nobody is in an office.

We are a small company and we write things down. Decisions are recorded, corrections are made to the source rather than to the output, and status is expected to be honest even when it is unwelcome.

Compensation and what we offer

• A long-term incentive, discussed at offer stage for senior appointments.
• Fully remote working, with company-provided equipment where the role requires it.
• A role with unusual scope for its level, because the organisation is being built rather than joined.
• Direct access to the founders. There are no layers between this role and the people making the decisions.

How we will assess you

We assess against three or four weighted signals agreed before we read a single CV, rather than against years of experience or a list of tools. Expect four stages:

• Introductory conversation with our People team. About 30 minutes, covering the role, your background and what you are looking for. This first conversation may be with an AI assistant; you will be told at the time and may ask to speak to a person instead.
• Conversation with the hiring manager. About 60 minutes, going deep on what you have done and how you think.
• A working session on a real ContactVerse problem. About 90 minutes. Not a whiteboard exercise and not a take-home test you do unpaid in your evenings.
• Final conversation with the founders. Covering fit, expectations and how we would work together.

We aim to complete the process within three weeks and to give you a decision, with reasons, within three working days of the final stage.

Equal opportunity

ContactVerse is an equal opportunity employer. We welcome applications regardless of age, disability, gender identity or expression, marital or civil partnership status, pregnancy or maternity, race, religion or belief, sex or sexual orientation. We select against the published signals for the role and nothing else. If you need an adjustment at any stage of the process, tell us and we will make it. It will not count against you.

Research consistently shows that some candidates apply only when they meet every criterion. If this role interests you and you meet most of what is listed, please apply.

How to apply

Write to career@contactverse.ai quoting reference POS-016, with your CV and a short note (a few paragraphs is plenty) on the part of this role you would most want to get your hands on, and why.