Case study · 2024 — Present
Mavik — Multi-Agent Orchestration & RAG Platform
Shared platform behind independent agentic products
Independent builder · Independent
The problem
The problem
Independent products needed a shared layer for agents, retrieval, and tools — not a one-off glue stack per app.
Outcomes
- RAG pipelines accelerating asset discovery ~10x vs prior workflows
- Multi-agent orchestration patterns reused across PlanEat, AgentOps, and internal tools
- MCP-style tool servers cutting custom integration work ~70%
Context
This platform spine sits under independent product work (maviklabs.com as the public studio site) — not under Launch Legends employment.
Retrieval quality, tool boundaries, and cost control matter more than demo chat UIs.
Approach
Separate indexing from query path; evaluate retrieval with golden queries.
Agents call typed tools (MCP servers) instead of ad-hoc function calling per feature.
Multi-model strategy by task (cost/latency/capability).
Key tradeoffs
Vector stores
Chose: Managed vector DBs by environment
Rejected: Single in-process store for all products
Operational isolation differs per product deployment.
Tool integration
Chose: MCP + typed contracts
Rejected: Per-feature glue
Permissions and reuse amortize across the studio.
What was hard
- Semantic failures that are not HTTP 500s
- Corpus/schema drift breaking retrieval quality
What I would change
- Earlier offline eval harness per corpus
- Stricter cost budgets per agent step
Artifacts
