Job Description

Company Description

We are Software Mind, an awesome team of engineers who are ready to ramp up any top-notch company’s projects! Our aim? To always be one step ahead. Become part of a multicultural company in constant growth with an excellent work environment certified by Great Place To Work!

Job Description

Project - the aim you'll have

Our client builds an AI copilot for process engineers in oil refineries and chemical plants: a natural-language interface where engineers ask questions about live plant data — equipment, sensor tags, process trends — and get grounded, chart-backed answers. The users are experienced engineers who are rightly skeptical of AI: in this domain, a fabricated number or a silent wrong assumption has real cost. The product wins or loses on whether the agent can be trusted.

This role owns the trust layer of that agent inside a large, active Python codebase. It is not feature work with an LLM endpoint bolted on. The work is the mechanics of agent reliability: making the agent say "I don't know" instead of inventing, surfacing every assumption it makes so the user can correct it, grounding every claim in actual data, holding output quality through model migrations, and keeping latency acceptable while doing all of the above.

To make the day-to-day concrete, this is what the engineer currently in this seat shipped in the last four months (all of it flag-gated, in small PRs, reviewed async daily by a team spread across the US and Australia):

• An assumption auditor: detects the silent assumptions the agent makes when answering (which equipment, which time window), validates them via multi-draw consensus, and surfaces them in the UI as correctable chips — the engineer can fix an assumption and rerun the analysis.
• A grounding auditor that catches reports fabricated from empty data feeds before they reach the user.
• An adversarial reviewer sidecar that critiques generated charts for correctness before display.
• Successive frontier-model evaluations (loop behavior, directive adherence, regression on a replay harness) that decided when to flip the product's default model — including, twice, deciding NOT to flip.
• Hardening of a plant-exploration tool against hallucinating structure that the data does not support.
• A latency fix: a narration side-loop was inflating query response times; capped it and made it best-effort.
• A concurrency fix making a shared data-reset path atomic, eliminating intermittent production read errors.

If reading that list is more interesting to you than building another CRUD feature, this role is for you.

Qualifications

Expectations - the experience you need

• Strong Python in large, shared, evolving backend codebases: you will work daily in code you didn't write, alongside people committing to it every day.
• You have shipped an LLM-based feature to production AND built an evaluation that changed a real decision (a model choice, a prompt rollback, a killed feature).
• Production debugging from symptom to confirmed root cause: latency spikes, concurrency errors, failures that produce no log line.
• Prompt work treated as engineering: measured adherence, regression testing against a fixed case set — not vibes.
• Comfort with feature-flag discipline and staged rollouts (default-off, soak, flip), small PRs, and mostly-async collaboration across US and Australia time zones.
• High autonomy: problems arrive ambiguous ("the agent feels slow", "the engineers don't trust the numbers") and you turn them into scoped, verifiable fixes without waiting for a spec.
• Direct, precise written English.

Nice to have

• GCP (Vertex AI in particular); AWS/Azure acceptable.
• Observability tooling (tracing, structured logging, latency percentiles you actually watched).
• Experience with charting/plotting pipelines (matplotlib or similar) feeding a UI.
• Industrial, process, or time-series data domain experience.
• Heavy AI-tooling development workflow (Claude Code or similar) — the team works this way.

What you will do

• Investigate agent misbehavior reported from real customer plants and turn each case into a diagnosis, a fix, and a regression test.
• Build and extend the evaluation harnesses that gate prompt changes and model migrations.
• Add reliability mechanisms to the agent: assumption surfacing, grounding checks, output-quality reviewers.
• Diagnose and resolve cross-cutting performance and concurrency issues.
• Raise code quality in the areas you touch, within the team's review conventions.

Our Benefits

• Educational resources
• Flexible schedule and Work From Anywhere
• Referral Program
• Supportive and chill atmosphere

We are accepting applications from LATAM countries