DwellFi

The architecture is agentic from the file system up, with agents writing custom software per fund against the customer's own stack rather than configuring a fixed product, and six composable primitives turning unstructured documents, email and agreements into structured tables agents can act on. Capital calls reconcile against the partnership agreement and net asset value strikes against custodian feeds, each tied back to source documents. Apply the removal test and what remains is the spreadsheets and headcount the company identifies as the status quo, since there is no configurable product underneath.

The human review loop is described as a mechanism rather than a principle, with exceptions caught by the customer's team feeding back to the agent operating in their environment so reviews get lighter over time, which implies review is the default and automation earns trust incrementally. Humans and agents share the same access model, and an audit trail covers every action. What is not published is the gate: no stated threshold at which an agent may complete a capital call or strike a valuation without sign off, and no sampling regime over work the agents did complete.

Three disclosed practices sit above the norm. Mathematical work compiles to code rather than being generated as text, and is cross validated across frontier models, which is a specific engineering answer to the arithmetic failure mode. Every answer ties back to a source document, so a reviewer can trace a figure to the agreement it came from.

And models are benchmarked on a continuous schedule with workloads routed accordingly, which is an ongoing evaluation regime rather than a one time selection. What is missing is published results: no accuracy figures, no benchmark outcomes, no model documentation and no validation package.

Two named fund administrators anchor the record and one is significant: a global provider selected the platform as its strategic partner for agentic fund operations, and a real estate fund administrator integrated the technology and markets itself as powered by DwellFi, which is a stronger commitment than a customer reference.

Team credentials are unusually deep for a company this size, including a former global chief executive of one of the world's largest independent fund administrators and engineering leadership from a major quantitative firm, with two patents held. Against that, total funding is modest at around five million dollars, no assets under administration figure is published, and no outcome is quantified at either named customer.

This is the most complete answer to the cross customer question anywhere in this index, and it is stated rather than implied: each agent learns from the work it does for a customer, contained to that customer's tenant, with the company stating plainly that nothing leaks across customers and that improvement is self contained within the perimeter. Every other network or learning vendor here either says nothing or describes anonymisation without a mechanism.

Two further disclosures support it: models are benchmarked continuously and each workload routed to whichever performs best, and mathematical work compiles to code cross validated across models rather than being generated as free text.

The privacy position is architectural and stated in absolute terms: the platform sits on top of the customer's existing stack and the company states that nothing moves and nothing leaves, with data, cloud and perimeter all belonging to the customer and inference running where they choose.

For a product handling limited partner records, side letters and capital account detail, keeping the data inside the perimeter the customer's auditors already accept removes most of the questions this index asks other vendors. An attestation is claimed. What is absent is documentation around retention and any published privacy framework.

A service organisation control attestation is claimed explicitly, described in one place as type one, and the security argument is reinforced structurally by running inside the perimeter the customer's auditors already accept, which means the institution's existing controls govern the environment. That is a stronger position than a hosted vendor asserting its own controls. What is absent is the surrounding evidence: no trust centre, no report request path, no stated audit period or scope, and no second framework.

DwellFi supplies technology and holds no licence, and its regulatory engagement is structural rather than rhetorical: the deployment surface is set by the customer's security review and regulatory regime before contract, with named regions including the European Union, United Kingdom, Singapore, United States and the Gulf states, so the platform adapts to whichever supervisor governs the fund.

Outputs land in regulated artifacts, since net asset value filings and investor reporting carry real obligations, and an audit trail on every action supports examination. No individual supervisory instrument is named as a design target.

The subjects are funds, documents and calculations rather than people, so this reads as accuracy governance, and the company sets the bar itself by stating that these are workflows where one hundred percent accuracy is the only acceptable outcome. Cross validation across models and compiling mathematics to code are real controls against silent error in a valuation.

What is absent is evidence that the bar is met: no accuracy figures, no error analysis by document or fund type, and no account of what happens when an agent's reconciliation and a fund accountant's disagree.

Two structural properties protect the customer more than most vendors here manage: the software runs inside their perimeter so they retain physical control of data and execution, and every output ties back to a source document so a disputed figure can be traced to the agreement behind it. Neither is a vendor commitment.

No accuracy guarantee accompanies the stated one hundred percent accuracy standard, no remediation term exists where an agent generated valuation feeds a wrong filing, and no route is described for a limited partner whose capital account was misstated.

The chain is described honestly in shape if not in names: the company states that it benchmarks frontier and open source models continuously and routes each workload to whichever performs best, which concedes plainly that external providers are in the path and that the selection is dynamic. Against that it claims every layer of the stack from parsing to reporting is its own, with no vendor between it and customer data, and inference running inside the customer perimeter. Individual model providers are not named and no subprocessor list is published, which is what holds this below the top grade.

The integration posture is additive by design and the company is explicit that no replatforming is required, with the layer sitting on top of the customer's existing ledger, customer system, documents and communications, reachable through an interface, a command line or a model context protocol server, which is the newest machine addressable integration standard and rare in this index. Turning every file, email and agreement into queryable structured tables is the harder half of the problem, and a fund administrator deployment reaches many underlying funds at once.

The strongest deployment position in the index. Software runs inside the customer's own environment, described as air gapped in some material, with data, cloud and perimeter all belonging to the customer and inference running where they specify. Residency is offered across named regions covering the European Union, United Kingdom, Singapore, United States and the Gulf states, and the location is fixed as a contract term rather than a vendor default.

For a product handling limited partner data across jurisdictions that is the correct architecture, and it exceeds even the on premise options elsewhere in this index because the choice is contractual rather than optional.

No rates, tiers, billing unit or minimum were located. One commercial term is disclosed and it is unusual and substantive: where the software runs is described as a contract term rather than a vendor preference, set by the customer's security review and regulatory regime before signing. That tells a buyer something real about how the relationship is structured without telling them the price.

Buyers span the private markets operations chain, covering fund administrators, private equity, real estate, venture and hedge funds, family offices and limited partners, with fund administrators the primary channel since one deployment reaches many underlying funds. Functional coverage is specific to the operational core, spanning capital calls, net asset value reconciliation, due diligence questionnaires, investor reporting and fund conversion. The boundary is that operational core: nothing addresses investment decisions, banking, lending, insurance or public markets.