Chief Scientist, Large Language Models

Role overview

One role, accountable for medical LLMs end to end

This is the company’s lead technical role for large models: you will own the full chain of medical LLM work — design, training, alignment, development and engineering deployment — and be responsible for building our general-purpose foundation capabilities.

Mission

Within three years: medical LLMs that are clinically usable, compliant to launch and proven by evidence

At the same time, you will build a general-purpose foundation model and an R&D organization that keep improving release after release. The role carries four mandates:

• Technical direction: Decide the foundation-model strategy — in-house pre-training, continued pre-training on open-source models, or a hybrid — and set the pace of investment in compute and data.
• Medical deployment: Bring models into real clinical care, hospital operations, medical insurance and drug and device R&D, and complete the required regulatory filings.
• Organization building: Build and lead a world-class LLM R&D team from zero to one.
• External influence: Represent the company to academia, industry, healthcare institutions and regulators.

Three-year roadmap

Build the foundation and complete filings, scale commercially, then set the industry benchmark

Three phases and two gates. Phase targets are planning goals and will be adjusted as the work progresses.

Responsibilities

Ten areas of responsibility

The Chief Scientist is accountable for the design, training, alignment, development, engineering and compliant launch of our medical LLMs.

1. 01Technical strategy and roadmap

   • Set a three-year technical roadmap for general-purpose and medical LLMs, including the foundation-model choice: in-house pre-training, continued pre-training of open-source models, distillation, and dense versus MoE architectures.
   • Assess frontier directions: reasoning models, reinforcement learning, multimodality, agents, long context and small on-device models.
   • Plan the budget and pacing of the three core resources: compute (GPUs and domestic Chinese accelerators), data and talent.

2. 02General-purpose foundation models

   • Lead pre-training and continued pre-training: data mixture, tokenizer, model architecture, scaling-law experiments and training stability.
   • Lead post-training: SFT, preference alignment (RLHF and DPO-style methods), reinforcement learning with verifiable rewards, and long chain-of-thought reasoning.
   • Keep the foundation model competitive on general capabilities — knowledge, reasoning, code, long context and tool use.

3. 03Medical LLM design and training

   • Build the medical model portfolio: medical Q&A, clinical decision support, medical-record generation and quality control, medication review, report interpretation, patient communication and medical-insurance cost control.
   • Build the medical corpus and knowledge system — textbooks, clinical guidelines, literature, drug labels, de-identified electronic medical records and medical knowledge graphs — with standards for cleaning, de-identification and quality.
   • Design evidence and safety mechanisms — source attribution, expression of uncertainty, refusal and referral policies — to reduce hallucination systematically.
   • Advance medical multimodality: joint modeling of medical imaging, pathology, ECG and laboratory data with text.

4. 04Alignment, fine-tuning and safety

   • Establish a physician-in-the-loop annotation, review and feedback cycle (RLHF with medical experts).
   • Build medical safety guardrails: interception of high-risk advice, red-teaming, bias and fairness evaluation, and defense against adversarial attacks.
   • Define standards for model interpretability and traceability.

5. 05Training and inference engineering

   • Lead distributed training at the scale of 1,000+ accelerators (Megatron-LM, DeepSpeed, FSDP, etc.), improving efficiency, stability and fault tolerance.
   • Lead inference optimization (vLLM, SGLang, TensorRT-LLM, quantization, speculative decoding, etc.) for cloud, private and on-premises hospital deployment.
   • Drive adaptation to domestic Chinese accelerators and China’s domestic IT stack (Xinchuang).
   • Build four platforms — data, training, evaluation and model serving (LLMOps).

6. 06Applications: agents, RAG and productization

   • Design the medical agent framework: tool use, multi-agent collaboration and integration with hospital information systems such as HIS, EMR, PACS and LIS.
   • Build medical retrieval-augmented generation (RAG) and knowledge bases covering guidelines, literature and in-hospital knowledge, with retrieval and citation.
   • Work with product and delivery teams to turn model capabilities into products that can be delivered at scale.

7. 07Evaluation

   • Establish two-tier evaluation — general and medical: public benchmarks (MedQA, MedMCQA, PubMedQA, CMB, CMExam, MedBench, HealthBench, etc.), in-house clinical-scenario evaluations and blinded physician review.
   • Put in place safeguards against data contamination, production monitoring and continuous iteration.

8. 08Compliance and regulation

   • Lead algorithm filing, launch filing and security assessment for generative AI services.
   • Where needed, drive registration of medical AI products as medical-device software with the NMPA.
   • Ensure compliance with China’s Interim Measures for the Management of Generative AI Services, the Data Security Law and the Personal Information Protection Law, including de-identification and secure management of medical data.

9. 09Team and organization

   • Build five teams — algorithms, data, engineering, evaluation and medicine — with clear talent standards and succession.
   • Establish R&D processes, experiment management, technical review and knowledge management.

10. 10External partnerships and influence

    • Run joint R&D and clinical validation with top-tier (Grade 3A) hospitals, medical schools and research institutes.
    • Publish at top conferences and journals, and contribute to industry standards and white papers.
    • Support fundraising roadshows and technical defenses for major clients and government projects.

Requirements

Five requirements

Education, publications, technical depth, healthcare knowledge and leadership all matter. Institutional background is a reference point for screening; achievements weigh more.

Preferred qualifications

None of the following is required; the more you bring, the stronger your candidacy

• Core author or technical lead on a well-known open-source LLM or medical LLM.
• A combined medicine and computing background (an undergraduate degree in clinical medicine or an MD, plus a PhD in computing).
• A best or outstanding paper award at a top international conference, or area chair at a top conference.
• IEEE or ACM Fellow, or selection for a national talent program.
• Led a medical AI product that obtained NMPA Class III medical-device registration or FDA authorization.
• Core R&D experience at a top industrial lab such as Google DeepMind, OpenAI, Meta or Microsoft Research.
• Entrepreneurial or successful commercialization experience, with an understanding of fundraising and capital markets.

Ideal candidates

Three profiles we are looking for

Our first choice is a candidate who combines industrial LLM training capability with a track record in medical AI. Next is Profile A. For Profile B, the company will pair you with a strong engineering deputy.

What we offer

Compensation and resources

Cash compensation, equity and signing bonus are determined in discussion, with no upper limit.

Compensation

RMB 2–5 millionAnnual cash compensation with no upper limit, plus equity/options and a signing bonus, determined in discussion.

Authority

Roadmap · team · budgetDecision rights over the technical roadmap, authority to build the team, and an annual R&D budget.

Compute

500PGPU and domestic-accelerator compute capacity, existing and planned.

Medical data

Nationwide dataMedical data and partnership resources with nationwide coverage.

Company stage

Angel roundRMB 80 million raised in the angel round.

Academic support

Joint appointmentsOpportunity for joint appointments with Tsinghua University, Xi’an Jiaotong University, Peking University, Harbin Institute of Technology and Sun Yat-sen University.

Assessment and process

Six assessment dimensions, eight steps

Technical depth carries the most weight; the process assesses technology, medicine, strategy and leadership together.

Assessment dimensions and weights

Interview process

1. Résumé screeningAssessment against the requirements and the self-check list.
2. Technical interviewAbout 90 minutes with the CTO or an external technical advisor, going deep into your technical work.
3. Research talkA one-hour presentation of your representative work, followed by Q&A.
4. Case presentationThe topic is shared one week in advance; you present a three-year medical LLM strategy on site.
5. Medical expert interviewExternal clinical experts assess your medical understanding and communication.
6. CEO / Chairman interviewStrategy, culture and values.
7. Background checkEducation, publications, reasons for leaving, references from former employers and peers, and non-compete review.
8. OfferOffer letter issued.

Before you apply

Check yourself against these eight points

If you do not meet a hard requirement, please explain why in your email.

1. A PhD, with an institutional background that meets the education requirement
2. Publications that meet the thresholds of either track (we verify author order and venues on Google Scholar and DBLP)
3. Led a complete training run of a 7B+ model through to production, and can state the data volume, compute scale, training duration and key metrics
4. A medical AI deployment, with the partner healthcare institutions and quantified results (or you meet the exception for medical experience)
5. Led an R&D team of 30 or more
6. No non-compete restriction, or one whose scope and term can be managed
7. Compensation expectations within this role’s range
8. Able to join full-time, with a clear start date

This role is probably not a fit if

• You contributed to, but did not lead, LLM training, and cannot explain the technical details under questioning.
• Your publications are mainly workshop papers, preprints or low-tier journals, or mostly in late author positions.
• Your role has been purely managerial, without hands-on technical work for years.
• Your experience is limited to application-layer work such as prompt engineering or RAG integration, without training or fine-tuning.
• You have any record of academic misconduct or data-compliance violations.
• You are seeking only a part-time or nominal role (advisory arrangements can be discussed separately).

How to apply

Please email the following materials to our recruiting contact

Contact

Zhang Xia

Email

echo@hxallai.com

Email subject

Application: Chief Scientist, LLMs – Your Name

Please include

1. Your CV (in English or Chinese)
2. A publication list marking formal acceptance vs. preprint; main conference vs. workshop or Findings; and first, co-first or corresponding authorship — with your Google Scholar link
3. A summary of one or two representative model projects: model size and architecture, training data (tokens), compute (accelerators × duration), key hyperparameters and lessons from failures, and core post-launch metrics
4. A summary of your medical AI deployment: partner institutions and quantified results
5. Your non-compete status and earliest start date

If you are bound by a non-compete agreement, please tell us its scope, term and compensation at first contact.

Apply by email