Home

TL;DR

Creator of OpenMed | Generative AI Leader in Healthcare | Building Sovereign On-Premise AI | Open Source AI Advocate

Product architect specialized in deploying state‑of‑the‑art Generative AI for regulated healthcare and life‑sciences settings, bridging public‑research rigor and enterprise production.

For over 16 years in public research, primarily at France’s National Centre for Scientific Research (CNRS), Maziyar has led large‑scale AI/ML platforms and model releases at the intersection of high‑performance computing, big data, and generative AI. He has been a core leader behind John Snow Labs’ Spark NLP ecosystem, which powers a vast share of medical NLP in production (including on AWS SageMaker and Amazon Bedrock). His stance is clear: medical AI must be open, auditable, sovereign, and deployable within a hospital’s own walls. He founded OpenMed to advance transparent, on‑prem‑ready medical language models.

Maziyar PANAHI: Full Professional Bio

Location: Paris, France

Principal AI Engineer and product architect specializing in bringing state-of-the-art Generative AI to the most demanding healthcare and life sciences environments. For over 14 years in public research, primarily with France’s National Centre for Scientific Research (CNRS), I have led large-scale projects at the intersection of high-performance computing and artificial intelligence. This deep scientific background informs my commercial work, where I led the team behind John Snow Labs’ Spark NLP, the most widely used NLP library in the enterprise, which powers the vast majority of medical AI models on platforms like AWS SageMaker and Bedrock.

My work is driven by a core belief: the future of medical AI must be open, auditable, and sovereign. This conviction led me to found OpenMed, an initiative dedicated to creating and sharing transparent, state-of-the-art medical language models. I believe that for AI to be truly trusted in clinical settings, it cannot be a black box. It must be deployable within a hospital’s own walls.

My expertise covers the full stack of medical AI deployment: from training and fine-tuning domain-specific LLMs (100M to 200B+ parameters) to architecting the national-scale, on-premise infrastructure they require. This includes managing billion-row databases on everything from bare-metal servers to hardened Kubernetes clusters. As an Infrastructure Project Manager at the Institute of Complex Systems of Paris (ISCPIF), I continue to manage the complex, high-performance computing environments essential for this pioneering research.

My mission is to leverage Europe’s leading AI to build the default engine for every hospital, pharma giant, and public health agency that requires a secure, sovereign, and powerful medical intelligence platform. I am focused on turning cutting-edge scientific research into robust, compliant, and commercially successful enterprise solutions.

Mission & stance

  • Build the default, trustworthy medical intelligence engine for hospitals, pharma, and public‑health agencies that require secure and sovereign AI.

  • Advance medical AI that is open, auditable, and deployable on‑prem, not a black box.

  • Turn cutting‑edge research into robust, compliant, and commercially successful solutions.

Current focus & roles (last ~7 years)

  • Leads and contributes to John Snow Labs’ Spark NLP ecosystem and related model releases.

  • Prolific publisher on Hugging Face with model cards, training configs, reproducibility notes, and evaluation summaries.

  • Hands‑on across post‑training and RL (SFT → preference → reward modeling → GRPO), evaluation harnesses, and red‑teaming.

  • Provides production‑grade artifacts and recipes for AWS SageMaker and Amazon Bedrock (incl. BYO Docker), plus tested inference configs for vLLM, TGI, SGLang, and llama.cpp.

  • Maintains curated families of general‑purpose (agentic) and medical models with consistent APIs and quantization matrices (2–8‑bit) for CPU edge, single‑GPU rigs, and high‑throughput servers.

  • Emphasizes “time‑to‑useful”: small, well‑documented defaults; safe fallbacks; tuning knobs for throughput/latency, context length, and tool‑use reliability.

OpenMed (founder): the open‑source, clinical‑grade alternative to closed, licensed products

What it is
OpenMed creates and shares transparent, state‑of‑the‑art medical LLMs and biomedical NER models that are free forever under Apache‑2.0, with reproducible training and transparent benchmarking. Built to run on‑prem or in your VPC, OpenMed is designed to integrate with hospital IT and regulatory workflows, so teams can ship HIPAA‑aware NLP and decision‑support without vendor lock‑in.

Scale & coverage
An expanding catalog of hundreds of models (475+ and growing) across 13+ biomedical categories, chemicals, diseases, genes/proteins, species, oncology, anatomy, and more, published on Hugging Face, backed by a Python package and CLI for one‑line pipelines.

Why it matters
Closed, licensed products keep clinical AI behind paywalls and black boxes. OpenMed is the state‑of‑the‑art open alternative: permissively licensed, auditable, and ready for sovereign deployment (on‑prem/VPC).

What OpenMed ships

  • OpenMed NER (SOTA): A suite of domain‑adapted transformers that achieve new state‑of‑the‑art on 10/12 public biomedical NER benchmarks, advancing micro‑F1 by up to +9.7 pp while remaining efficient to train(LoRA + DAPT) and easy to deploy.
  • Production toolkit: openmed Python package & CLI with: curated model registry, one‑line pipeline creation, advanced NER post‑processing, formatting (dict/JSON/HTML/CSV), input validation, and de‑identification helpers for clinical text workflows.
  • Deployment recipes: AWS SageMaker marketplace packages & JumpStart notebooks for five‑minute endpoints; Docker images & guidance for on‑prem/VPC with observability, encryption, and audit logging.
  • Extensibility: Lightweight LoRA adapters, curated tokenizers, and starter notebooks to extend entity coverage to local ontologies and multilingual records at modest compute cost.

Differentiators vs. closed‑source

  • Price & access: Apache‑2.0 licensing; no per‑seat/volume fees; self‑host anywhere.
  • Auditability & trust: Reproducible training, detailed model cards, transparent datasets; designed to help teams meet emerging regulatory expectations (e.g., EU AI Act) with private deployment and audit trails.
  • Sovereignty: On‑prem & VPC‑first posture; no mandatory call‑outs to third‑party APIs.
  • Performance: Peer‑reviewed results with SOTA on broad benchmarks; competitive with, and often exceeding, closed, licensed systems.

Call to action
Explore the models and Model Discovery app on Hugging Face, deploy a SageMaker endpoint in minutes, or install the Python package to add clinical‑gradeLLM to your pipelines.

Public‑research career (CNRS / ISC‑PIF)

Institute of Complex Systems – Paris Île‑de‑France (ISC‑PIF, CNRS), Dec 2015 → Present (lifetime civil servant)

Roles: AI Platform Leader; Principal Research Computing Architect; Infrastructure Project Manager; information‑security responsibilities.

  • Architected and operated national‑scale AI/ML platforms supporting distributed NLP/LLM/GenAI workloads (TensorFlow, PyTorch, ONNX, Apache Spark).

  • Built multi‑cluster big‑data architecture handling:

    • 360B+ records on Hadoop/Spark (Cloudera)

    • 7B+ documents in Elasticsearch

    • 16B+ records in MongoDB

  • Ran 140+ servers (≈2,000 cores, ~320 TB storage), including ~280 TB HDFS.

  • Deployed across AWS, Azure, and private clouds (OpenStack, OpenNebula).

  • Led the open‑source Multivac platform (multivacplatform.org) and contributed to large data/intelligence programs (e.g., Tweetoscope, Politoscope, Journalist).

  • Drove end‑to‑end infrastructure, security compliance, and DevOps practices for research at scale.

Selected impact & proof points

  • Open LLM Leaderboard (v1 & v2): Top performer from launch through archival; final placements include #1 and #2.

  • Quantization at scale: Published thousands of GGUF/GPTQ/AWQ variants with consistent metadata, licensing, checksums/signatures; tuned for CPU edge, single‑GPU, and multi‑GPU servers.

  • Adoption: Fine‑tuned general‑purpose agentic and medical LLMs used across research and industry.

  • Production recipes: Reproducible deployments for Hugging Face, AWS SageMaker, and Amazon Bedrock (including Bedrock‑ready artifacts and BYO Docker for SageMaker).

  • Release hygiene & trust: Benchmark‑first model cards; versioned collections for easy checkpoint selection by task, size, and hardware budget.

Technical expertise

Healthcare NLP & LLMs

  • Spark NLP leadership, medical VLMs, and medical reasoning LLMs.

  • Building and evaluating domain‑specific models (100M → 200B+ parameters), with emphasis on safety and clinical utility.

Distributed ML & Engineering

  • JVM/Spark production ML; Databricks integration; large‑model training/fine‑tuning; evaluation pipelines; observability for inference.

  • Inference stacks and optimizations for vLLM, TGI, SGLang, llama.cpp.

Big‑data platforms & infrastructure

  • Architecture and ops for on‑prem + cloud; hardened Kubernetes; data‑at‑scale systems; security/compliance for regulated contexts.

  • Multi‑cloud deployments (AWS, Azure, OpenStack, OpenNebula) and hybrid environments.

Post‑training & RL

  • Full post‑training pipelines (SFT → preference → reward → GRPO loops), eval harnesses, and red‑teaming.

  • Publishing best‑practice artifacts (configs, cards, metrics) for reproducibility and auditability.

Model publishing & tooling

  • Standardized metadata, versioning, and signatures; consistent APIs; quantization matrices (1–8‑bit).

  • CI’d exporters and packaging for Bedrock/SageMaker and common inference backends.

Representative releases & projects

  • Medical VLM: multimodal medical vision‑language model release on AWS SageMaker and BedRock.

  • Medical reasoning LLMs: domain‑focused models with safety‑aware evaluation and red‑teaming.

  • Spark NLP: multi‑year leadership across releases, how‑tos, and enterprise integration patterns.

  • Quantization families: curated GGUF/GPTQ/AWQ variants with side‑by‑side evals and deployment configs.

  • Production recipes: reproducible artifacts and guides for AWS SageMaker (incl. BYO Docker) and Amazon Bedrock.

Speaking & community

  • Frequent speaker at the NLP Summit and related industry events.

  • Contributor to the Databricks blog (e.g., scaling ViTs with Spark NLP).

  • Maintains a large Hugging Face footprint with public model cards, training configs, and evaluations.

Ideal Roles

  • Founding GM, Healthcare & Life Sciences (research → product → GTM → P&L)

    Charter
    Build and scale a sovereign medical AI product line (LLMs/VLMs) that is open, auditable, and deployable on‑prem or in VPC, with clinical‑grade safety and operational reliability.

    What I own

    • Product: Vision, strategy, and roadmap; packaging (on‑prem/VPC, managed, SDKs); pricing & licensing; design‑partner program; solution playbooks per ICP.
    • Research & Models: Data governance; SFT → preference → reward → GRPO; evaluation harnesses & red‑teaming; quantization matrix (2–8‑bit); inference stacks (vLLM, TGI, SGLang, llama.cpp).
    • Engineering & Platform: APIs, SDKs, agents, connectors; deployment targets (hardened Kubernetes, Bedrock, SageMaker, air‑gapped); observability, SLOs, rollback & canary.
    • Safety, Security & Compliance: PHI handling and GDPR/HIPAA alignment; audit trails, model cards, DSRs; responsible‑AI gates; security controls toward SOC2/ISO‑27001 as needed.
    • GTM: ICP & segmentation (providers, pharma R&D, CROs, payers); messaging & positioning; sales playbooks; field enablement; alliances (AWS, Databricks, EHR vendors); marketplace listings.
    • P&L & Ops: Budget, hiring, vendor management; OKRs; pricing/margin; partner programs.

    Proof from OpenMed & CNRS

    • OpenMed: reproducible model releases; top leaderboard placements (#1, #2); thousands of quantizations; production‑ready artifacts for Bedrock/SageMaker and common inference backends.
    • CNRS/ISC‑PIF: architecture & ops for national‑scale platforms (360B+ records; 140+ servers; multi‑cloud), with security and reliability in sensitive environments.

  • Head of Generative AI for Healthcare (own models, evals, and safety for clinical use).

  • Principal Applied ML (NLP/LLM) with mandate to ship on Spark/JVM stacks.

  • ML Platform/Infra Lead for regulated, data‑intensive orgs (on‑prem + cloud, security).

Search keywords

Maziyar Panahi, Spark NLP, OpenMed, medical LLM, medical VLM, GRPO, RLHF, quantization (GGUF/GPTQ/AWQ), vLLM, TGI, SGLang, llama.cpp, Bedrock, SageMaker, CNRS, ISC‑PIF, Multivac Platform.

Also, you can find me here: