gabe pereyra

Cursor's Composer product demonstrates the end-state Harvey aims for: packaging synthetic data, Neo lab partnerships, and post-training into a served model that runs alongside closed-source models in production.

N-gram is working with Harvey on enterprise search and firm knowledge capabilities, part of the multi-lab partnership strategy to build domain-specific AI infrastructure.

Coda's horizontal, individual-focused productivity tools (Coda Work) cannot address the organizational workflow needs of large law firms — coordinating 20-30 person teams across 6-month projects, resource allocation across thousands of matters, and multi-firm enterprise coordination — which is Harvey's vertical moat.

Trajectory provides post-training expertise and infrastructure for training NVIDIA's NeMo-Megatron open-source models, contributing to Harvey's strategy of leveraging multiple Neo labs for different model architectures.

Applied Compute is collaborating with Harvey on the Vault product, likely addressing secure deployment of post-trained models for enterprise customers with sensitive data requirements.

Snorkel's platform helps Harvey scale the synthetic data generation process initiated by domain experts, building larger training datasets for post-training legal AI models.

LangChain's tooling helps Harvey make expensive RL evaluation environments efficient, critical for running thousands of LLM-as-judge unit tests on large diligence datasets during post-training.

Fireworks provides critical infrastructure for post-training open-source models and serving them in production, enabling Harvey to train models like GLM 5.1 with advisor models and integrate open-source models into their serving stack alongside closed-source providers.

Mercor enables Harvey to scale domain-expert-guided synthetic data generation, solving the data privacy bottleneck for vertical AI by allowing lawyers to create realistic training datasets without using confidential client data.

Harvey's playbook proves application-layer companies can achieve frontier-level intelligence on specific domains by combining domain-expert synthetic data (via Mercor/Snorkel), multi-lab post-training partnerships (Fireworks, Baseten, Trajectory), and production-grade model serving infrastructure with routing/fallbacks — all without frontier-lab-scale capital.