Member of Technical Staff, Research Engineer

Member of Technical Staff, Research Engineer — Models, Training, Evaluation, Inference

Type: Full-time

About Hussh

Hussh is an AI-first research, machine-learning, and product organization. We are building Private Agent One: a private personal AI agent — personal superintelligence that works for one human, and one human only. The human owns their information, controls it, and shares it only by explicit consent, through our PCHP (Personal Consent Handshake Protocol). That constraint — real intelligence, running for a person, under that person's control, on their devices and on our cloud — is the research problem. It is not a smaller version of the frontier problem. In several ways it is the harder one.

The lab works from Kirkland, WA, with research partnerships at IIT Kharagpur and Stony Brook's AI3 institute. We also run a permanent-capital asset-management business (Fund A). We are founder-funded, and the research agenda is set by the work, not by a publication calendar or an advertiser.

The role

You build the intelligence inside Private Agent One: the models, the training pipelines, the evaluation systems that keep us honest, and the inference stack that serves a personal agent fast, cheaply, and privately — on device and in the cloud.

This is a research engineering role in the frontier-lab sense: you run experiments end to end, you build the infrastructure your experiments run on, and you ship what works into a product a person trusts with their whole information life. A result that never reaches the product is a hobby. A product claim our evals can't back is a lie. You own both sides of that line.

What you will do

• Design, train, and improve the models behind Private Agent One: fine-tuning, post-training, and adaptation of frontier and open models to personal-agent tasks — memory, retrieval, tool use, planning, and action under consent.

• Build and own the evaluation stack: task benchmarks, regression suites, human-evaluation design, and automated judges calibrated against human judgment. Your evals decide what ships.

• Own inference end to end: serving architecture, latency and cost budgets, quantization and distillation for on-device models, and the hybrid routing between device and cloud.

• Build the data engine: curation, filtering, deduplication, synthetic data generation, and the feedback loops that turn consented product usage into better models.

• Work on relevance as a first-class problem: retrieval, ranking, and recommendation over a person's own information — the ranking function serves the user, and only the user.

• Publish internally with rigor: experiment reports a skeptical reader can reproduce. Externally, where the work serves the mission and the lab's partnerships.

Expectations

• Empirical first. Hypotheses are cheap; you design the experiment that kills or confirms them, and you report what the data said — especially when it wasn't what you hoped.

• Velocity with reliability: you make researchers around you faster, and your training and eval infrastructure is infrastructure other people can depend on.

• You ship to product. You will sit with the engineers and designers who turn your model into an experience, and you will care whether it worked for the human.

• Taste counts as a technical skill here: which problems are worth solving, which results are real, which demos are lying.

Required qualifications

• 10+ years of professional experience in machine learning, research engineering, or closely related software engineering, with production systems to show for it.

• The last 5 years building AI products — shipped models in real products, not only papers or prototypes — with depth in at least one of: personal information systems, subscriptions, ads, or relevance/ranking.

• Deep, current fluency in modern ML: transformer architectures, training dynamics, fine-tuning (SFT), post-training and alignment methods (RLHF/RLAIF, DPO, RL methods such as PPO/GRPO), and PEFT techniques (LoRA and successors).

• Large-scale training experience: distributed training across many GPUs, with the debugging scars to prove it — parallelism strategies, checkpointing, failure recovery, throughput optimization.

• A record of evaluation work: you have built benchmarks or eval pipelines that real decisions depended on, and you can explain how you validated the eval itself.

• Bachelor's degree in computer science, mathematics, statistics, physics, or a related field — or equivalent demonstrated research-engineering achievement. Advanced degree welcome, never a substitute for shipped work.

Technical requirements

• Languages and frameworks: Python at expert level; PyTorch required, JAX a strong plus; CUDA/Triton for performance work; comfort reading and modifying framework internals.

• Training stack: DeepSpeed, FSDP, or Megatron-LM-class distributed training; Ray for orchestration; experiment tracking with Weights & Biases or equivalent; Hugging Face ecosystem fluency.

• Inference stack: vLLM, TensorRT-LLM, or SGLang in production; quantization (GPTQ/AWQ-class methods); KV-cache and latency optimization; on-device inference with llama.cpp, MLX, Core ML, or ExecuTorch.

• Data stack: large-scale data pipelines (Spark or equivalent), tokenization, quality filtering and deduplication at corpus scale; embedding systems and ANN retrieval (FAISS/ScaNN-class).

• Infrastructure: GPU clusters on AWS, GCP, or Azure, or Slurm-managed clusters; Docker and Kubernetes; profiling with Nsight/torch profiler; observability for training and serving.

• Ranking and relevance: learning-to-rank, embedding-based retrieval, and online experimentation (A/B design, interleaving, guardrail metrics) — from systems where ranking quality was measured against real user outcomes.

Compensation

Base salary: $280,000–$420,000 USD per year, plus a founding equity grant (4-year vest). Where you land in the range depends on experience and demonstrated results. Founding titles are permanent: no future re-leveling takes yours away.

Benefits

Comprehensive health, dental, and vision coverage. A 401(k) plan. Employer-matched contributions to Trump Accounts. Support for continuing education, retooling, and retraining — including AI courses, which we actively encourage. Above all: you will work with the latest AI technologies every day, and the expectation — for engineers and non-engineers alike — is to become expert in using them, running evals on your own work until it reaches an A+ standard.

Benefits are subject to the terms of the applicable plans and to eligibility requirements, in accordance with federal and state law.

We read every note ourselves.

Hussh is an equal opportunity employer. We hire on merit, competence, and qualifications, and do not discriminate on the basis of race, color, religion, sex, sexual orientation, gender identity, national origin, age, veteran status, disability, or any other legally protected characteristic.

1. Describe a research result you turned into a shipped system — the idea, what broke in production, and the evaluation that proved the shipped version was better.

2. Tell us about an evaluation you designed that changed a decision — the metric, the baseline, and what the result made the team do differently.

3. Give us a failure: a model or method that looked right in the lab and failed on real users — the number that exposed it, and what you changed.