// Careers

Production-ML consultancy in Wrocław.
Currently not hiring.

bards.ai is a small senior team. We open a role when a specific person would materially improve what we ship. If you want to be in the queue when that happens, this page tells you what we work on, what we look for, and how the process runs.

6+ yrsmedian tenure of seniors
16+open-source models
100M+prod tokens / day (largest customer)

// What we work on

Recent customer work and OSS.

Engagements run 2-12 weeks. Most are embedded with a customer's eng lead. We scope, ship, and stay on for the eval rollout.

  • Brand24

    Slack-native agent unifying 13 internal data sources / 8M+ datapoints, sub-5s p50 latency. We built the orchestration, eval harness, and on-call observability stack.

    LangGraphPydantic-AILangfuseQdrant

  • SurferSEO

    Production LLM pipeline processing 100M+ tokens/day. Custom fine-tunes for SEO content, prompt cascades for cost, structured outputs throughout.

    vLLMLoRAOpenAI / AnthropicStructured outputs

  • Comcast

    UI-element detector running daily on millions of VOD screenshots. Pre-annotation tooling that replaced a manual QA loop. Multi-config model routing.

    YOLOv8Ray ServeTritonONNX

  • Open source

    16+ models on huggingface.co/bardsai. ~80K downloads/month. Whisper-PL, Polish twitter sentiment, finance-sentiment in 7 languages, Jaskier (DPO Mistral-7B).

    DPOQLoRAGGUFSpeech / NLP

// Stack

What you'd touch.

We pick the stack per engagement. Below is what comes up most. You don't need fluency in everything; you need fluency in deciding.

  • LLM / agents

    LangGraph, Pydantic-AI, DSPy. OpenAI, Anthropic, Gemini, self-hosted. Structured outputs (Instructor, Outlines, GBNF). Langfuse + OpenTelemetry traces.

  • Training / fine-tuning

    PyTorch, transformers, Axolotl, TRL. LoRA / QLoRA, DPO, KTO, GRPO. Synthetic data pipelines. Eval harnesses (lm-eval, custom).

  • Inference / serving

    vLLM, TGI, Triton, Ray Serve. ONNX / TensorRT-LLM where it pays off. Continuous batching, paged attention, speculative decoding.

  • Retrieval / data

    Qdrant, pgvector. Hybrid search + rerankers (Cohere, BGE). Document AI / OCR (LayoutLMv3, Surya). Postgres, ClickHouse, S3-compatibles.

  • Languages / infra

    Python is the default. Some TypeScript at the edges. Go and Rust where latency matters. Docker, K8s, Terraform. AWS / GCP / on-prem.

// Bar

What we look for.

Not a checklist - a description of the people on the team. If three or four of these clearly fit you, apply.

  • 01

    5+ years shipping ML or LLM systems to production - or equivalent research depth with a track record of getting models out of notebooks. We hire papers-and-prod people; not papers-only, not prod-only.

  • 02

    Can argue for or against fine-tuning vs prompt engineering vs retrieval given a specific problem, with cost and latency numbers. "It depends" answers should come with the dependencies.

  • 03

    Read a flame graph, debug a tool-call trace, and write the postmortem. We do our own on-call.

  • 04

    Comfortable owning the boundary: data → model → infra → eval → customer call. We don't hand off between roles.

  • 05

    Wrocław weekly is the default. Remote across EU works for the right person; we don't do US-overlap roles.

// Process

Four steps. Roughly two to three weeks end-to-end.

0130 min · video

Intro

What you've shipped, what you're working on now, whether the scope here matches. Either side can pass at the end with no hard feelings.

0290 min · video

Technical deep-dive

Pick a project you led end-to-end and walk us through it. We dig into trade-offs you made, what broke in prod, and what you'd change. No leetcode, no whiteboard.

033-4 hours · paid

Paid pair-work

Anonymized problem from our queue: an eval design, a debugging session, or a model-selection brief. Live screenshare or take-home, your call. Paid at our standard contractor rate.

0460 min · video or in person

Founders

Comp, scope, start date, what setup makes this work for you. We answer everything we can.

// FAQ

The questions we get asked most.

  • What's the comp range?

    Senior engineer / staff equivalents. We pay at the top of the Polish market and adjust for EU remote when relevant. We share a band before pair-work; expect details, not theatre.

  • Hybrid, remote, in-office?

    Wrocław weekly is the default for engineering. Remote across EU timezones for senior IC roles when the work fits. We don't run a US-overlap rotation.

  • Do you hire juniors / new grads?

    No. The work is customer-facing from day one and the team is small. We open associate roles only when a specific senior wants to mentor someone full-time.

  • Are you open to contract instead of full-time?

    Yes, for engagements of 3+ months. Same bar, same rate band as our full-time IC level.

  • I'm not ready to apply but want to stay in touch.

    Email hello@bards.ai with a one-liner about your current work. We keep a list and reach out when a role opens that fits.

// Apply / get in queue

Email hello@bards.ai.

Reply within one business day. We read everything; we don't run an ATS.

What to send

  • - Code we can read: GitHub, repos, gists. Anonymized prod snippets are fine.
  • - Papers, OSS, or shipped product you'd point to.
  • - One paragraph: what you're working on now and what you'd want next.
  • - CV optional. LinkedIn fine if your repo is private.