Person
1 episodes · 23 quoted claims
Aman Singh leads Arca's AI platform team, which builds the harness used to run the business with AI.
“my role at Arca is I lead the AI platform team, so we're building the underlying sort of, um, harness and our capabilities to kind of run the business with AI”
Arca offers a broad range of financial services including financial advice and planning, with tax, insurance and estate planning.
“this is a broad spectrum of financial services, so that includes, um, sort of financial advice, financial planning for, um, for their assets”
Arca moved its primary model provider from Gemini to OpenAI and then to Anthropic over the years.
“we used to be in Gemini way back when, um, and then OpenAI and then and then Anthropic.”
Arca found that a model called Luna, praised in external benchmarks, underperformed in its own tests.
“despite the external benchmarks claiming, um, great things for Luna, it didn't really quite work out for us as as well as we had expected.”
Arca uses an internally built harness that is model-agnostic instead of provider-specific harnesses.
“we have our, our own, uh, kind of built internally that is model model agnostic.”
Arca assigns its heaviest tier of models to tasks such as tax planning and tax analysis.
“a task like assisting with a tax plan or doing a tax analysis would get our, like, heaviest tier of models.”
In Singh's tax-bill example, agents analyze how best to liquidate a client's holdings to raise $100,000.
“what is the most optimal way, um, to liquidate some of that to to produce $100,000?”
Arca's agents cannot automatically move money or liquidate assets.
“they're unable to actually execute, uh, any sort of money movement or liquidation automatically.”
Arca has a deep integration with its custodian that provides APIs for money movement and liquidation.
“We have a very deep integration with our custodian, uh, that actually gives us APIs to be able to do that.”
Arca keeps persistent memory for each client covering current status, ownership, history and next steps.
“for every client, um, what is happening right now, what are we trying to service, who owns what, what has happened so far, what are we tracking for next, right?”
Arca runs LLM-based verification on every tool call that writes data.
“with every time a tool is called that writes data, we have, um, uh, LLM-based verification running on, uh, on that action.”
Arca uses a consensus of small models to check outgoing data for exfiltration attempts.
“but also a consensus of small models that that ensure that there isn't any kind of data exfil being attempted.”
Every agent action at Arca produces a change log recording what changed, who changed it, from which part of the system, and when.
“every single action that that an agent takes produces a change log of what changed, um, who changed it, from which part of the system, and when.”
Arca is wary of depending on Anthropic's managed agents product.
“but less so on on taking a dependency on a product like managed agents.”
At Arca the sandbox is one component of the harness, and the agent uses it to offload computation instead of running inside it.
“the sandbox is one component of our harness. Our agent isn't actually inside of the sandbox. It uses a sandbox to offload computation.”
Arca's verifiers check generated financial documents for non-compliant content such as promised returns.
“you really should not be, like, um, promising returns, uh, in in that artifact.”
Singh says models may confidently use outdated tax brackets, such as those from 2025, in their analysis.
“models will very happily use the tax brackets from 2025, um, to to come up with an analysis”
Arca does not currently use agents for autonomous trading; agents only propose trades.
“at the moment, we are not doing autonomous trading with with agents.”
Arca runs a nightly process that collects real-world agent traces and uses a meta agent to suggest harness changes.
“every night we run this process where we take collect traces from, uh, the real world from our agents in the wild and run, run a meta agent on top that suggests changes to our harness.”
Arca still requires a human engineer to approve prompt changes before they ship.
“today, we still have a a human in in like a human engineer in the loop before we actually ship a prompt change.”
Arca's product has a Knowledge tab holding the operational context agents use, which experts can edit.
“on the left hand side, there's a tab called Knowledge, which has the full, uh, operational context that our agents are using.”
In Singh's example, a prospective client said their existing advisor charged $1,500.
“They're charging me $1,500, um, uh, for for their services.”
Singh says Arca's agents proactively flag client risks such as an approaching deadline or a missed required minimum distribution (RMD).
“someone has a deadline coming up, or they're at risk of missing their RMD”