Founding Research Engineer

Client: AI trading technology company (confidential)

Location: Remote

Reports to: Chief Executive Officer

Type: Full-time. Senior / staff-level individual contributor with a path to leading the research function.

The role

I’m working with an AI trading company building an agentic intelligence and execution layer that sits on top of broker and exchange infrastructure. Their agents plan, size, execute and manage positions for real users with real money, on live venues.

Frontier models are excellent generalists, but none of them has seen how retail traders actually behave: how they size, when they override, where they blow up, and what a well-managed position looks like against a risk budget. The client has access to that data through their own agent trajectories and their partners, and they are going to post-train models on it.

This is the first hire to do that. You would own the benchmark, the data, the training runs, the serving stack and, most importantly, the decision on whether the client’s own model beats the frontier on their tasks.

What you would own

• The evaluation. Design and build the client’s benchmark of agentic trading tasks: execution workflows, risk-budget adherence, position sizing, instrument grounding, multi-step tool use across venues, and decision explanation. It becomes the standard every model is scored against, theirs or a frontier API. The evaluation is the product; the model is downstream of it.
• The data. Turn live agent trajectories, user overrides, outcomes and partner behavioural data into training-ready datasets, with anonymisation, consent and data residency handled correctly from day one.
• The models. Run supervised fine-tuning, preference optimisation and reinforcement learning on open-weight base models in the 30-70B class, scored against the benchmark, with an explicit kill-or-scale decision at each stage.
• Serving. Get the winning model into production at acceptable latency and cost, behind the same interfaces their API-served models use today.
• The decision. Tell the CEO, with numbers, whether the model beats the frontier on their tasks. If it does not, explain why and what would change that.

What they need

• Post-training shipped on an open-weight LLM that reached production or a published result: supervised fine-tuning and at least one of DPO, GRPO or PPO-style RL. A LoRA in a notebook does not count.
• An evaluation harness that other people depended on, and an understanding of why the evaluation comes before the model.
• Experience training models for agentic behaviour, such as multi-turn trajectories, tool and function calling and environment interaction, or clearly equivalent work.
• Competence on the data side: pipelines, deduplication, filtering, PII handling, synthetic data generation and its failure modes.
• Production serving experience (vLLM, SGLang, TensorRT-LLM or equivalent) and the ability to reason about quantisation, latency and cost.
• Strong Python and PyTorch; fluent with the Hugging Face stack (TRL or similar); comfortable with distributed training on rented GPUs (FSDP, DeepSpeed).
• Judgement. Knowing when not to train, and willing to kill your own project on the numbers and say so plainly.
• Able to operate alone for the first stretch, inside a regulated financial business, and write clearly.

Nice to have

• Built RL environments or simulators
• Market microstructure, trading or risk-management literacy; exposure to retail brokerage or crypto derivatives
• Time-series or quantitative background (useful for features and evaluation, not the core of the role)
• Publications or open-source contributions in post-training or evaluations
• Experience with data residency and regulated data across multiple jurisdictions

Interview process

• Deep dive on one fine-tune you shipped: data, method, evaluation, what moved, what did not, and what you would do differently.
• Design exercise. Design the evaluation for one agentic trading task and a post-training plan against it.
• Founder conversation covering judgement, the kill decision, and how you work without a team around you.
• References focused on shipping, not credentials.

Why join

• Greenfield. You define the research function, the benchmark and the training programme from nothing.
• Proprietary data. Live agent and broker partner data nobody else has, with the rights to train on it.
• Real deployment. Your model executes with real money on live venues, not in a demo.
• Direct line to the CEO. No layers.