Grounded Agents (stealth, backed by Greycroft) provides an AI executive assistant that maps a leader's professional world (people, projects, dependencies, decisions) and proactively surfaces what needs their attention. As the founding engineer I built the system end to end:
- LLM provider layer. Routing across Anthropic, Google, and OpenAI models over Bedrock, Vertex, and native APIs, with a bidirectional circuit breaker that fails over between platforms and a fallback selector that matches context-window size rather than capability tier alone. Also reasoning-effort control and a model catalog.
- Context and cost control for long-running agents. Input truncation, token budgeting, and compaction so hour-long runs land a result instead of getting cut off; per-run turn caps set from p99 production telemetry; prompt caching holding ~70% hit rate.
- Durable agent runtime. Designed and implemented an agent execution pipeline based on Temporal workflows, providing resumable runs, scheduled fan-out, and turn and token budgets with an explicit wrap-up path.
- Knowledge graph (Neo4j). Provenance-first: every claim points at a message or document a person actually wrote, so a model's inference can't be promoted into a stated fact. Entity resolution reconciles the same person, project, or decision across Slack, email, and documents into a single node.
- Observability and spend attribution. Metric taxonomy, log classification, and the dashboards and monitors for the LLM path: per-model, per-agent, per-tenant cost attribution and failover visibility.
- BYOC (bring your own cloud). Customer-deployable stack based on Terraform, an umbrella Helm chart, and a reconciler that keeps a customer-run installation in sync. Plus the required security infrastructure: Teleport just-in-time access, WAF, TLS enforcement on the graph store, KMS.
|