Allia
Allia brings AI into accounting workflows where documents arrive at volume, domain context matters, and every answer needs to remain useful under real operating constraints.
What I worked on
- Designed the multi-agent runtime with LangChain and LangGraph, hybrid retrieval, domain tools, and MCP integrations.
- Built evaluation and tracing with Langfuse to measure grounding, citations, latency, cost, and failure modes.
- Made high-volume document ingestion durable with Temporal workers, bounded concurrency, and batched vector writes.
In production
- Approximately 340,000 workflows each week.
- Approximately 89% fewer critical failures in internal evaluation.
System: Agent harnesses · Context engineering · MCP integrations · Temporal · Langfuse
The recurring lesson is simple: the harness and feedback loop matter as much as the model. Context, tools, evaluation, and runtime behavior must work as one system.