April 28, 2026 — TMF Group has chosen DwellFi for agentic fund operations. Read more

Platform

The architecture behind the OS.

Three principles compile to one architecture.

[Sovereign

Deploy where and how you want. The customer's call, set at contract.](/content/platform#sovereign/index.html) [Agentic

Every fund is one of one. Real-time software, per fund, written by deterministic agents.](/content/platform#agentic/index.html) [OS

Composable building blocks. Deterministic execution across six core primitives.](/content/platform#os/index.html)

Sovereign.

Agentic.

OS.

AI TABLE · LP COMMITMENTS

47 rows

Westbrook Growth Fund III

CITED FROM 47 LPAS · CROSS-VALIDATED

Sovereign

Deploy where and how you want.

Where DwellFi runs is a contract term, not a vendor preference. The customer's security review and regulatory regime set the deployment surface before signing, and the OS runs inside the perimeter the customer's auditors already accept.

The institutional pattern is the same one DwellFi ships: software runs inside the perimeter the customer's auditors already accept. AI workloads run inside that same perimeter.

Deployment surfaces

Your own data center

or air-gapped

Region of your choice

Geography

EU, UK, Singapore, US, GCC, or the region your regulator requires. Data residency and inference both stay there.

01Infrastructure

AWS to

air-gapped.

AWS, Azure, GCP, your own data center, or an air-gapped enclave. You pick the surface. The OS deploys to it.

DEPLOYMENT SURFACE

CLOUD

ON-PREMISE

AIR-GAPPED

CUSTOMER-DEFINED

02Region

The region your

regulator requires.

EU, UK, Singapore, US, GCC. Data residency and inference both stay in the region your regulator requires.

DATA RESIDENCY

EU

UK

SG

US

GCC

DATA STAYS IN-REGION

03Tenancy

Single tenant,

your account.

Your DwellFi instance runs inside your cloud account. No shared infrastructure. No cross-tenant compute or data.

TENANCY MODEL

CUSTOMER 01

CUSTOMER 02

CUSTOMER 03

ONE TENANT PER CUSTOMER

Agent · Capital call

Run 4127 · In-tenant

Westbrook Growth Fund III

capital call · 47 LPs ·$24.2M

Execution trace

01 →Read LPA §4.2 (commitment terms)

02 →Pull LP commitments from AI table

03 →Calculate pro-rata:14.5% × commit

04 →Cross-validate (Claude, GPT-4o, Gemini)

05 →Draft notices · 47 LP-specific letters

06 →Awaiting reviewer

Deterministic

3 of 3 models agree

Audit

2026-05-23 09:14:02Z

Deployed inside the customer’s tenant, against this fund’s LPA. Every action lands in the audit log.

Agentic

Every fund is one of one.

No two funds share the same combination of side letters, asset mix, jurisdictional exposure, regulatory regime, waterfall mechanics, and valuation policy. This is why vertical SaaS falls short in fund operations, and why the work still runs on Excel and headcount.

DwellFi writes real-time custom software, per fund, against your stack. The agents compose the workflow. Capital calls reconcile against the LPA. NAV strikes against custodian feeds. Each workflow ties back to its source documents the same way, every period.

Generic AI guesses. DwellFi agents generate, verify, and ship.

Inside agentic · Orchestration

The best model for each task.

DwellFi benchmarks frontier and open-source models on a continuous schedule, then routes each workload to whichever performs best on it. The fund isn’t locked into one lab’s roadmap.

For workflows that have to be exact, deterministic agents cross-check across multiple frontier models and route exceptions to a human for review.

Model pool

Anthropic

OpenAI

Google

Meta

Mistral AI

Routed by continuous benchmark. No single-lab lock-in.

The agents above are the work. The OS below is what makes them deterministic.

01Composition

Agents compose

the workflow.

A capital call agent reads the LPA, queries the commitments table, drafts the notice, and hands off to the review queue. The composition is what makes it a workflow. Not a prompt, not a chat thread.

AGENT WORKFLOW

READ

QUERY

DRAFT

HANDOFF

COMPOSED AT RUNTIME

02Determinism

Math becomes

code.

Pro-rata math and waterfall logic become deterministic code. Same inputs produce the same numbers, every run. Six models verify before any number reaches a human.

DETERMINISTIC EXECUTION

RUN 01

$12,847,332.18

RUN 02

$12,847,332.18

RUN 03

$12,847,332.18

SAME INPUT, SAME OUTPUT

03Per fund

Each fund gets

its own software.

Side letters and fund-specific waterfall provisions. The agents write the code that handles them, per fund, against your stack. Configuration was the old answer. Per-fund code is the new one.

FUND-SPECIFIC CODE

FUND 01

FUND 02

FUND 03

PER FUND, NOT SHARED

OS

An OS that scales the work, not the headcount.

Agents are only as good as the layer they run on. The DwellFi OS is that layer: composable primitives that turn fund documents into deterministic answers, every one tied back to a source.

Primitives

5 objects · Infinite compositions

01

Org Members

Humans and agents share the same access model.

02

AI-Native File System

Files understood by AI on day one. Indexed in place.

03

AI Tables

Structured rows extracted from documents, cited line by line.

04

AI Agents

Deterministic workers trained on fund operations.

05

AI Skills

Codified institutional processes. Composable across agents.

capital calls, NAV closes, allocations, K-1s. different compositions of the same primitives, not different products.

Architecture stack

Top-down · Customer in

Layer 1

Access

Chat portal · Developer console

Layer 2

Agents and Skills

The services customers use

Layer 3

Primitives and Integrations

AI-Native File System · AI Tables · MCP

Layer 4

LLM-agnostic

Claude · GPT · Gemini · Llama · DeepSeek · Mistral

Layer 5

Cloud-agnostic infrastructure

AWS · GCP · Azure · Oracle · your data center

Layer 1 is what the customer touches. Layer 5 is where it deploys. The middle is the harness.

Probabilistic vs deterministic

Same input. Different throughput. Different output.

Generic AI is a model that produces text. Fund operations are reconciliation problems, not text problems. Four real workloads, four different ways probabilistic systems fall over.

Generic AI

Probabilistic

Throughput

1. 01 →Reading PDF
2. 02 →Processing pages 1–32
3. 03 →context_length_exceeded after token 8192

Output

Unable to process file. Try splitting into smaller documents.

Deterministic

Throughput

1. 01 →Indexed 120 pages
2. 02 →Identified 47 LP sections
3. 03 →Extracted side letter terms per LP
4. 04 →Cross-validated across six models

Output

47 LP side letter records. Every term cited to its page.

Generic AI

Probabilistic

Throughput

1. 01 →Parsing statements
2. 02 →Recognized 12 of 36 months
3. 03 →Pattern inference: skipped 24 months

Output

Partial reconciliation. Confidence: low. 24 months not processed.

Deterministic

Throughput

1. 01 →Parsed 36 monthly statements
2. 02 →Three-way match against custodian and GL
3. 03 →Identified 412 breaks, every one categorized

Output

36-month reconciliation complete. 412 breaks, each with a reason.

Generic AI

Probabilistic

Throughput

1. 01 →Uploading 250 files
2. 02 →Request payload too large
3. 03 →Connection timed out at 600s

Output

File set exceeds API limits. No output.

Deterministic

Throughput

1. 01 →Queued 250 documents, 49,832 pages
2. 02 →Streamed through KYC and sub doc pipelines
3. 03 →Generated 250 LP profiles, every doc cited

Output

250 LP profiles complete. AML and KYC files attached.

Generic AI

Probabilistic

Throughput

1. 01 →Opening file
2. 02 →Read tab 1: Summary
3. 03 →Read tabs 2–5
4. 04 →Output truncated · 25 tabs not processed

Output

Partial extraction. 25 tabs and all cross-sheet formulas dropped.

Deterministic

Throughput

1. 01 →Parsed 30 tabs
2. 02 →Normalized to a 24-column schema
3. 03 →Validated formulas and cross-sheet references

Output

All 30 tabs in one normalized dataset. Every formula traced to its source cell.

Same input. Different system.

Workflows

Everything that has to be right.

The work that scales with AUA or AUM, not headcount. The work that has to reconcile, not approximate.

Capital calls

Notices drafted from LPA commitments, sent to LPs of record, reconciled against custodian receipts.

NAV close

Positions, accruals, fees, and waterfall calculations rolled into a deterministic NAV pack.

Reconciliations

Bank, custodian, broker, and ledger reconciliations against source feeds.

Distributions

Calculated, drafted, approved, executed against the waterfall.

Investor reporting

Quarterly LP letters, capital account statements, side-letter compliance.

K-1 generation

Tax allocations, capital accounts, K-1 packages per LP.

Investor onboarding

Subscription document review, KYC, AML, accreditation, side-letter review.

Form PF, Form ADV

Regulatory filings, drafted from current fund state.

Fee and waterfall

Management fees, performance fees, carry, hurdle, catch-up, waterfall.

Audit prep

Trial balances, supporting schedules, audit confirmation packages.

Fund migrations

Inbound migration of an existing fund admin's data into the OS.

Custom workflows

Per-fund agents written against your stack.

The security floor for institutional buyers.

SOC 2 Type II, with audit trails stamped on every action. Deployment runs inside the customer's environment. The full security overview covers the controls and the deployment architecture.

Read the full security overview → Request the SOC 2 Type II report →

See the OS handle your fund operations.

Sovereign by design, deterministic by construction, auditable on every action, deployed inside the customer's environment. Pick the conversation that fits where you are.

See specific use cases → Talk to sales →