Braintrust

Researched 2026-06-28. Primary category: llm-observability; also ai-red-teaming (eval/scoring of model outputs).

One-liner — Eval-first observability platform for production AI: log traces, score outputs, turn failures into test cases, and gate releases on evals.

What it does — Braintrust unifies trace logging, automated evaluations (LLM-as-a-judge and custom code scorers), prompt management, and dataset curation into one workflow. The signature loop: capture every prompt/tool call in production, score outputs, one-click-convert failures into test cases, and gate CI/CD releases on eval results. Built around a custom database for AI observability data and tuned for long-running, multi-step agents.

Where it sits in the stack — Primary home is the llm-observability layer; its scoring/red-teaming-style evals also touch ai-red-teaming. Dev/CI- and ops-time tooling, not an inline runtime firewall — it is not an inline prompt/egress control, though eval gates can catch unsafe behavior pre-release. Complements ai-runtime-security.

Deployment & architecture — SDK-based instrumentation, SaaS by default with self-hosting for enterprise. Deep CI/CD integration for eval gating is a core design point. Framework-agnostic.

Positioning & differentiators — Known as the eval-first / “make evals a first-class CI gate” platform, contrasted with observability-first tools. Versus langfuse and arize-phoenix it is closed-source and leads with evals rather than OSS tracing; versus langsmith it is framework-neutral and eval-centric; versus helicone it is SDK/eval-based not proxy-based. Strong logos (Notion, Stripe, Vercel, Replit, Ramp, Cloudflare) signal traction with AI-native engineering teams.

Ownership, funding & M&A — Independent, VC-backed. Series B $80M announced 2026-02-17, led by ICONIQ (a16z, Greylock, Elad Gil, basecase participating); press reports ~$800M valuation. Prior $36M Series A (Oct 2024, led by a16z, ~$150M post). Founded 2023 by Ankur Goyal (ex-SingleStore/MemSQL, ex-Figma); HQ San Francisco. No M&A. Confidence: high on round facts; valuation is press-reported (med).

CTO / hedge-fund lens — Day-1 if your bar is “no AI ships without passing evals.” The CI-gating model fits a fund that wants a promotion gate between experiment and production (ties to promotion-gates). Self-hosting addresses keeping eval data (potential MNPI) in-boundary. Well-funded, so continuity risk is low. Not a governance/SR 11-7 platform by itself — pair with ai-governance-platform.

Competitors / alternatives — langfuse, langsmith, arize-phoenix, helicone, datadog, comet, patronus-ai, maxim-ai.

Open questions / to verify

• Confirm the ~$800M Series B valuation against a primary disclosure (press-reported).
• Self-hosted feature parity vs SaaS.

Sources

• Braintrust’s Series B (company blog) — fetched 2026-06-28 — supports: $80M Series B, ICONIQ lead, 2026-02-17, positioning, customers; confidence: high (valuation med — press).

History

• [2026-06-28] Stub created from seed registry.
• [2026-06-28] Researched; confirmed independent (Series B $80M Feb 2026, ICONIQ), filled founding/HQ/funding/deployment; kept dual category (llm-observability + ai-red-teaming). No M&A. ownership_confidence low→high.