What tech stack does Decagon use?

Decagon's backend is built around a proprietary multi-model AI stack. As of March 2026, approximately 80% of inference traffic runs on in-house fine-tuned open-source models — trained specifically on enterprise customer support conversations — rather than third-party APIs from OpenAI or Anthropic. The architecture uses an 'ecosystem of agents' model where multiple specialized agents work together and review each other's outputs rather than routing all queries through a single model.

The serving layer uses Together AI's inference infrastructure, running on NVIDIA HGX B200 (Blackwell) GPUs with high tensor parallelism and speculative decoding via custom draft models. Together AI's end-to-end latency was selected over other options specifically for voice-speed workloads; production achieves P95 model latency under 400ms per turn even on multi-thousand-token inputs, while reducing inference cost by approximately 6× versus closed models. Prompt caching is also employed to further reduce per-turn costs at high volume.

Cloud infrastructure spans Google Cloud (primary, with Google VPC and Cloudflare WAF for network security), Microsoft Azure AI Foundry (used to deploy fine-tuned models close to users for latency and data residency compliance), and AWS US-East (referenced in production infrastructure contexts). The multi-cloud posture is intentional, enabling Decagon to satisfy enterprise data-residency requirements across US and European regions. Decagon holds SOC 2 Type II certification, is GDPR-compliant, and offers HIPAA-eligible configurations for healthcare clients.

Decagon's frontend is built in React and TypeScript, confirmed by job postings for Frontend Engineers specifying React, TypeScript, HTML/CSS, and RESTful API experience. The company's internal quality and observability infrastructure includes Watchtower — a proprietary always-on QA monitoring system that flags live agent conversations requiring human review based on configurable criteria — and Agent Workbench, a debugging and performance visualization tool launched in early 2026 that autonomously identifies and surfaces issues in production agent behavior.

On the product integration side, Decagon connects natively with Salesforce (full Customer 360, case creation and update, opportunity stage modification), Zendesk (two-way ticketing sync, knowledge base search, escalation routing), Shopify, Stripe, ServiceNow, and Confluence. These are customer-facing integrations built into the product itself — not necessarily the internal GTM and sales tools Decagon's own revenue team uses, which have not been publicly disclosed. The internal toolset for Decagon's go-to-market function is not confirmed from public sources.

Decagon's deliberate shift to in-house models signals a strong build-over-buy posture for core AI infrastructure. Vendors pitching LLM API access or off-the-shelf AI model layers will face resistance unless they offer specific capabilities — such as specialized voice latency, multimodal understanding, or domain-specific fine-tuning — that Decagon's in-house stack cannot yet match cost-effectively. By contrast, vendors in the infrastructure layer — GPU compute, model observability, data pipelines, security and compliance tooling — have strong natural fit given the scale of Decagon's inference workloads.

Decagon's multi-cloud posture (GCP + Azure + AWS) means the company is not locked to any single cloud provider's marketplace. Direct sales motions are necessary rather than relying on marketplace co-sell. Security and compliance vendors (data residency, SIEM, identity and access management) have a strong pitch given that Decagon must satisfy SOC 2 Type II, GDPR, and HIPAA-eligible requirements for financial services and healthcare customers. Any vendor already serving Decagon's enterprise customers (Chime, Rippling, Avis Budget Group, Deutsche Telekom) can use those reference relationships as a credible entry point for conversations with Decagon's engineering and operations teams.