Software Engineering for AI Labs
Software engineering for AI labs and data vendors: RL environments built from production code, SWE tasks with hidden verifiers, and the engineers behind them.
EltexSoft works with AI labs and AI training data vendors on two things: reinforcement learning environments built from production codebases, with hidden verifiers, Harbor and HUD formats and reward-hacking review, and the engineers who build them, as a pod under one contract or as single roles. Pilots are fixed price and graded by the buyer. We do not name labs or vendors on this site.
What we ship
The Work
-
RL environments from production code
A Dockerized production codebase with its database, cache and workers inside, running with the network off. Tasks mined from merged pull requests, each with a hidden verifier and a gold solution, delivered in Harbor and HUD formats.
-
Expert coding data
SWE tasks written and graded by senior engineers who have maintained the kind of codebase the tasks come from. Instructions a model can act on, tests it cannot see, difficulty labels from measured model runs.
-
Reward-hacking review
An engineer who did not write the task tries to pass the verifier without doing the work. Pass-to-pass tests make partial fixes score zero. Tasks with a shortcut go back to the author before a model sees them.
-
Calibration against real models
Every task is run against a weak and a strong model before delivery. Tasks that every model solves or no model solves are cut; the rest are labelled easy, medium or hard.
-
The engineers behind the environments
Task authors, verifier engineers, environment infrastructure and QA and calibration, contracted as a pod under one agreement or as single roles on your own environment team.
-
Contract terms built for labs
NDA before any detail is shared, IP assignment on delivered work, identity and location verification for every engineer with project access, project-only access with synthetic data, professional liability insurance.
AI labs and the vendors that supply them buy two things from engineering companies: environments a trainer can trust, and the people who can build more of them. EltexSoft offers both, built on the same senior engineers who have shipped production software for clients in healthcare, legal, eCommerce and AI for over a decade.
RL environments
Our environments start from a production codebase, ours or one you supply, frozen at a snapshot and packaged as one Docker image with the database, cache, background workers and stubs for external APIs inside. Tasks are mined from merged pull requests reverse-applied onto that snapshot, so each one is a change a human engineer actually made. Every task carries a hidden verifier (fail-to-pass tests that must turn green, pass-to-pass tests that must stay green), a gold solution and a reward rule, and ships in Harbor and HUD formats. The full validator, grading rules and pilot terms are on the RL environments page.
The engineers
Four roles make an environment: task authors, verifier engineers, environment infrastructure and QA and calibration. The person who writes a task never grades it. The roles are available as a pod under one EltexSoft contract, or one at a time on staff augmentation terms for labs that already run their own environment program. Details on the hire RL engineers page.
What we do not do
We do not publish task counts or pass rates from client work, we do not name labs or vendors on this site, and we do not quote rates on the site. Ask, and an engineer replies within one business day.
FAQ
Common questions
What does EltexSoft build for AI labs?
Do you work with data vendors as well as labs?
How does a pilot work?
What are your rates for AI lab work?
Tell us what you're building.
One business day reply. From an engineer, not a sales rep.