Founders · Vancouver, BC

Built by two founders in Vancouver, BC.

AutoStack was founded by Riya Arora and Bhavya Bakshi in Vancouver, BC. Bhavya leads engineering, including the environments and graders; Riya leads product development and operations. AutoStack builds environments for training and testing AI agents, starting with an insurance claims environment.

01 · Product development and operations

Riya Arora

Co-founder, product development and operations

Riya leads product development and operations at AutoStack: what each environment covers, which workflows it tests, how an engagement runs from the first call to the readout, and how we bring it to market.

02 · Engineering

Bhavya Bakshi

Co-founder, engineering

Bhavya leads engineering at AutoStack: the environments, the MCP tool servers, seeding and resets, and the deterministic graders that score every episode.

Why environments

Why we build environments for AI agents

AI agents are starting to act inside business systems: recording decisions, setting reserves, issuing payments, deciding when a human has to step in. Most testing still reads what an agent says about its work. The expensive mistakes are in what it actually did.

So we build working copies of those systems, with written rules and graders that check the final state and every attempted action. Labs train on them. Teams test on them before their customers find the failures.

agent_said
→ escalated to supervisor
audit_log
→ payments_issue rejected

What it said vs did

environment
→ seeded state
→ tools over MCP
→ graded on change

What we build

So far

What we've built so far

2026-10Three insurance clientsThree insurance clients so far, across pilot and paid engagements. Client names stay confidential.
liveThe insurance claims environmentA fictional insurer, Larkspur Mutual: 15 tools over MCP, a claims manual with rule IDs, and a public practice set of 15 tasks on 3 seeds each. Scored evaluations use private tasks. See it →
2026-10-08The grader, trustedDeterministic grading from the final state and the audit log, checked against do-nothing and shortcut agents.
2026-10-10The baseline runThree models, 45 episodes each, on the 15-task public practice set. Read the labelled results →

How we work

How we work with customers

01One engagement at a timeSo your workflow gets our full attention, from the first call to the readout.
02Synthetic data onlyWe never need your production data. Security and data →
03Written permissionWe test an agent only after its owner has asked us to, in writing.
04Plain answersEmail riya@autostack.ca, or request an evaluation and we reply within one business day.

Talk to the people who build it.

Request an evaluation