About WebLife
WebLife is a pioneering force in the eCommerce industry, with over $250 million in all-time revenue and a 16-year track record of building and scaling successful online businesses. As the largest distributor of mailboxes across the United States, WebLife has built a reputation for operational excellence, customer focus, and digital innovation.
Centrifuse is WebLife's AI venture building FridayOS, an opinionated operating system built on organisational memory and AI agents. This is a high-energy founding team that combines the creative speed of a startup with the stability and resources of an established company.
Backed by WebLife's long-term stability and operational maturity, we operate with the speed and ownership mindset of a product-first startup — tight feedback loops, high accountability, and an expectation that every hire meaningfully influences strategy, execution, and how our ventures evolve over time.
Combining the strength of a proven business with the agility of an innovation hub, we offer a collaborative, growth-focused environment where your ideas matter and your career can thrive.
About the Role
You will be directly shaping how organisations capture what they know and run their work through AI. You will build the pipeline that turns the traces our agents already produce into small open-weight models that replace the narrow, repeated LLM calls at a fraction of the cost, run on a customer's own desktop, and never carry one organisation's data into another's. You are the founding engineer for this.
The first year's outcome is a product-level set of distilled models behind a measured gate, and the capture, consent and isolation design that lets per-organisation adapters follow.
What You'll Do
- Trace capture. Every LLM call stored as a trainable record with its context, output, feedback, outcome, organisation and consent flag, with retention and deletion rules that hold up when a customer leaves.
- The eval gate. A model quality score with a cost and latency axis, built on the existing test suite's cost tracking and benchmark, read before any model is promoted. You own the number everyone trusts.
- Distillation. Frontier-model outputs filtered by the eval into training sets; supervised fine-tuning and preference tuning of small open models (Qwen3, Llama 3.x, Gemma 3 class) with LoRA on rented GPUs; shadow deployment alongside the frontier model before promotion.
- Serving. Tuned models behind the existing provider abstraction, vLLM in the cloud and Ollama on the desktop, with adapter versioning and rollback.
- Per-organisation adapters. Isolated training, storage and serving per organisation, retrain-from-scratch on deletion, and an exportable bundle of training set, eval suite and adapter that the organisation owns.
- The decision not to train. A compiled context pack per organisation is the cheaper first answer, and you will prove where it is enough and where weights win.
What We're Looking For
- 5+ years of solid software engineering experience, with 2+ years building production LLM systems.
- Hands-on fine-tuning of open-weight models: LoRA and QLoRA, supervised fine-tuning and DPO, with Unsloth, Axolotl or Hugging Face TRL, and the judgement to know when a smaller base model is the right choice.
- Distillation or synthetic-data work with a frontier teacher: generating, filtering and curating training sets from a stronger model's outputs.
- Building evaluation harnesses for LLM outputs, including LLM-as-judge with calibration against human labels, and using them as a release gate rather than a report.
- Serving open models in production with vLLM or a comparable engine, including multi-adapter serving and quantisation trade-offs.
- Data pipeline work: turning traces and feedback into versioned, reproducible training sets with provenance, consent and deletion handled by design.
- Comfort with structured-output tasks (classification, extraction, constrained drafting), since that is most of the work.
- Pragmatic product sense: you ship the narrow model that saves money this quarter before the general model that might.
Nice to Have
- Multi-tenant ML experience where tenant isolation was a hard requirement.
- Experience with an agent framework such as LangGraph, or with building agent systems directly.
- Familiarity with OpenTelemetry-based tracing of LLM calls.
- Rust or TypeScript, since the product is Rust and Svelte, and the provider layer is where your models land.
- Open-weight licence literacy (Apache 2.0, Llama community licence, Gemma terms) and the habit of recording it next to the model.
- Experience working with US, UK, European or Australian teams or clients.
Why Join Us?
- Competitive salary with payment in USD
- Work from the comfort of your home in a fully remote setup
- Be part of a culturally diverse, inclusive, and innovative team
- Opportunities for professional development and growth within the company
- Founding-engineer ownership of a new product capability
