AI Trainer

DevFixr; Pune District; On-site; INR 1,653,000 - 3,306,000; Part time

Job summary

DevFixr seeks a software task designer to craft high-quality RL-ready engineering tasks. You will patch real code, produce unambiguous instructions and tests, and ensure outputs have a single correct interpretation.

The role emphasizes hands-on coding depth across Python, with broader language flexibility, and requires clear written English. Immediate start in a sprint-driven, hourly engagement is available.

Qualifications

• Real engineering depth from shipping and maintaining production systems.
• Years on CV matter less than judgment; some strong contributors are very young.
• Python preferred; language-agnostic: Java, C++ or COBOL experience valuable for tasks moving legacy code to modern languages.
• Intellectual honesty: raise problems early; say you’re nearly done only when truly close to completion.
• Clear written English; most output is written specification readable without questions.

Responsibilities

• Write realistic software engineering tasks by inspecting repos, changing or fixing code, applying patches, and cleaning up afterward.
• Write precise instructions, test cases and expected outputs so there is exactly one correct interpretation.
• Run frontier models against tasks and evaluate outputs against strict functional/logical standards.
• Contribute to RL environments for languages beyond Python where needed.

Job description

Engagement

Contract, paid hourly. No 40-hour minimum — work arrives in sprints, and you take on whatever suits you.

Start

Immediately. The client has sample work that needs turning around this week, with a substantially larger programme expected to follow.

About the work

Our client builds high-fidelity datasets used to train and evaluate frontier large language models. One of the things the AI labs want most right now is software engineering tasks for reinforcement learning environments — specifically, tasks that frontier models cannot solve.

The RL environments are already built. What the client needs is a steady supply of tasks to run inside them, written by engineers who know what hard, real work actually looks like.

What you’ll be doing

• Writing realistic software engineering tasks: go into a repository, change or fix something, apply a patch, clean up afterwards. The emphasis is on work an engineer would genuinely be paid to do not competition-style puzzles or textbook exercises.
• Writing the instructions, test cases and expected outputs tightly enough that there is exactly one correct interpretation.
• Running frontier models against your tasks and evaluating the responses against rigorous functional and logical standards.
• Where your expertise sits outside Python, helping build out the RL environment for that language.

The bar

A good task defeats a frontier model on the merits: the model understood exactly what was being asked and still could not do it.

A task that "wins" because the wording was loose does not count. If a model returns something you weren’t looking for because you didn’t specify properly, that is a fault in the task, not in the model.

Holding that line — hard, but scrupulously unambiguous — is most of the job.

What the client is looking for

• Real engineering depth. The kind that comes from shipping and maintaining production systems.
• Years on a CV matter less than the quality of your judgement; some of the client’s strongest contributors are very young.
• Python, or a very good reason not to. Python is the primary environment. But the client is language agnostic: if you have spent a career in Java, C++ or COBOL and never picked up Python, that expertise is genuinely valuable — there is real demand for tasks that move legacy code into modern languages.
• Intellectual honesty. You say so when you don’t know something, you raise problems while they are still fixable, and when you say you’re nearly done, you’re nearly done. This matters more than almost anything else.
• Clear written English. Nearly everything you produce is written specification that someone else has to be able to read without asking you a question.

Helpful, but not required

• Prior work on AI data platforms — Outlier, Alignerr, Scale, Surge or similar RL and data-annotation workflows.
• Experience across several codebases, stacks or domains rather than one.

Requirements

What the client is looking for

• Real engineering depth. The kind that comes from shipping and maintaining production systems.
• Years on a CV matter less than the quality of your judgement; some of the client’s strongest contributors are very young.
• Python, or a very good reason not to. Python is the primary environment. But the client is language agnostic: if you have spent a career in Java, C++ or COBOL and never picked up Python, that expertise is genuinely valuable — there is real demand for tasks that move legacy code into modern languages.
• Intellectual honesty. You say so when you don’t know something, you raise problems while they are still fixable, and when you say you’re nearly done, you’re nearly done. This matters more than almost anything else.
• Clear written English. Nearly everything you produce is written specification that someone else has to be able to read without asking you a question.