Agentic knowledge acquisition studio

Internal control plane tying multi-agent task graphs, retrieval, and evaluator hooks for regulated dossiers (stub brief).

Cold-start agent swarm <11 min median (staging)

Role: Director of AI · architecture + implementation
Stack: LLM gateways · Postgres · OpenSearch · async workers · policy-as-code guards

Context

BillionHands-aligned scenario: SMEs must ingest heterogenous dossiers weekly. Manual copy/paste dominates; hallucinated merge proposals are unacceptable for downstream compliance tooling.

Approach

Decompose dossier ingestion into specialist agents with explicit memory tiers, deterministic tool contracts, evaluator agents that rerun golden datasets hourly, and a CI-like promotion flow from sandbox to prod clusters.

Outcomes

From greenfield connector to swarm online in < 11 minutes median in hardened staging VPCs — mostly Terraform + IaC templating.
Evaluator hooks reduced tool misuse incidents −41% during dogfood week (classification by safety bot).

Agents
Orchestration