round3
round3 builds the tools for self-owned AI: a secure, observable, developer-first stack for specialized language models that organizations can build from their own data, align with their values, and run on their own terms. Its first product, three.dev, finds failures in live AI-agent traffic and helps coding agents fix them.
At a Glance
- Software teams shipping AI features or agents in production
- Developers and product teams that need evidence for model and prompt changes
- Companies with high-volume or latency-sensitive LLM workflows
- Organizations seeking to own specialized language models
- +1 more
AI Tools by round3
(1)three.dev
AI Agent Failure Detection
Discussions
No discussions yet
Be the first to start a discussion about round3
Latest News
three.dev is now generally available, with automatic live-traffic failure discovery and coding-agent fixes.
Published a 2,000-request comparison of seven frontier models on PII redaction across quality, latency, and cost, showing that the benchmark winner was not the best fit for the use case.
Published research showing that a nominal model price cut can increase the actual bill because of tokenization, output mix, caching, and reasoning behavior.
Published a production case study on detecting and fixing an AI support-assistant identity-verification failure with three.dev and its MCP integration.
Products & Services
A production AI quality and observability platform. Teams route LLM calls through its proxy; it records and scores sampled live traffic, identifies and ranks recurring failure modes with evidence, supports comparison of configurations before and after changes, and gives coding agents the context to fix failures. round3 also uses it to help teams evaluate and build specialized language models for particular workflows.
Market Position
three.dev positions itself as a production-traffic-first alternative to evaluation and observability tools. Its published comparisons distinguish it from Braintrust (datasets, scorers, and CI), Langfuse (open-source tracing and configured evals), and LangSmith (broader agent lifecycle tooling): three.dev emphasizes automatically naming failure modes from live traffic with little setup, then handing evidence to a coding agent for remediation. It also differentiates through quality/latency/cost comparisons and a path from production data to specialized models.
Leadership
Founders
Borja Burgos
Co-founder and CEO. Previously co-founded Tutum, a platform for running Docker applications acquired by Docker in 2015; co-founded Undefined Labs, acquired by Datadog in 2020; and led product for developer-focused observability at Datadog.
Ximo Guanter
Co-founder and CTO (LinkedIn identifies him as Joaquín Guanter Gonzálbez). Previously a software developer at Microsoft, Senior Principal Software Engineer at Telefónica, co-founder and CTO of peer-to-peer Bitcoin exchange Coinffeine, and Staff Software Engineer at Datadog.
Executive Team
Borja Burgos
Co-founder & CEO
Third-time startup founder; co-founded Tutum and Undefined Labs, both acquired, and previously led developer-focused observability product at Datadog.
Ximo Guanter
Co-founder & CTO
Previously worked at Microsoft and Telefónica, co-founded Coinffeine as CTO, and was a Staff Software Engineer at Datadog building large-scale data systems.
Founding Story
The founders saw that teams could change prompts, models, or agents quickly but had difficulty knowing whether a change actually improved their product in real production traffic. round3 was started to replace production-AI decision-making based on demos and generic benchmarks with evidence from each team's users, constraints, and data, ultimately enabling teams to create and own specialized AI.
Business Model
Revenue Model
Subscription SaaS priced primarily by plan, stored data, scored requests, and history; enterprise customers receive custom limits. All plans include scoring, failure modes, coding-agent access over MCP, chat, and unlimited seats.
Pricing Tiers
About 50,000 requests/month at 5k tokens each; 1 GB data/month, 5,000 scored requests/month, and 30 days of history.
About 250,000 requests/month at 5k tokens each; 5 GB data/month then $3/GB, 10,000 scored requests/month then $2.50/1,000, and 90 days of history.
About 1,000,000 requests/month at 5k tokens each; 20 GB data/month then $2/GB, 25,000 scored requests/month then $1.50/1,000, and 180 days of history.
High-volume plans with custom data, scoring, and history limits sized to traffic.
Target Markets
- Software teams shipping AI features or agents in production
- Developers and product teams that need evidence for model and prompt changes
- Companies with high-volume or latency-sensitive LLM workflows
- Organizations seeking to own specialized language models
- Language learning, customer support, and other domain-specific AI applications
- Monitoring and improving customer-facing AI agents in production
- Finding recurring failures in support assistants and other conversational agents
- Evaluating model, prompt, tool, and reasoning changes against real traffic
- Choosing an AI configuration across quality, latency, and cost constraints
- Building specialized small language models for a narrowly defined business workflow
- AI features such as language correction, ticket triage, summarization, and other domain-specific tasks
- LanguaTalk: used three.dev to compare 40+ variants for its Spanish-corrections feature and build a specialized language model that matched target quality at 282ms p50 and 56% lower cost than Sonnet 4.6.