A98 is an operating system for private market funds: credit, venture, PE and family offices. Analysts and executives design their own agents and pipelines on the firm’s data, inside the firm’s own cloud, without an engineering team. Two rules underneath. The AI reads and decides what to check, but never does the math. And nothing gets overwritten.
Concept demos, built to test the idea, on fictional portfolios: the ask, the steps, the numbers computed by code, what gets flagged instead of guessed, and the report it fills at the end. Click a type to switch.
Four fund types in the concept: private lending, private equity, venture capital, family office. The ontology is defined per firm, not per strategy, so the same machinery handles a borrower’s covenant and a portfolio company’s ownership. If your firm uses a metric, your instance computes it.
A98 is not a fixed menu of features. Your firm defines its own ontology, its own source hierarchy and its own rules, and your team builds the agents and pipelines on top. Another fund gets a different instance, because it is a different fund.
Your data room, your agreements, your closed quarters. A98 turns them into sourced facts, every figure traced to a page, and resolves which version governs when sources disagree. Every deduction shows its evidence.
Before anything goes live, the system re-derives figures you already trust from quarters you already closed: a covenant test, a fund IRR, a distribution waterfall. It goes live only when its numbers match yours.
Agents and pipelines your team builds, organized by your teams, writing to your own warehouse of results. The fund owns the data, the ontology and everything it builds, and can take all of it with it.
Two building blocks, in plain terms, and what they do to your team's week.
An agent reads the exact documents for one task, computes with deterministic code, and returns the answer with its sources. It can be recurring (the quarterly covenant check) or a one-off you’ll never run again (“what’s our exposure to this supplier that just went under?”). Your team assembles one from the firm’s own pieces, in plain language or in the structured builder, and it produces the same object either way. You can also point one at a single detail you care about, and it watches that detail every cycle.
A pipeline chains agents and runs itself, monthly or quarterly, without anyone asking. Every run is versioned into your firm's warehouse, and the last step doesn't produce a chat answer: it fills the reports you already use. The credit memo, the LP letter, the board pack, your template, your numbers, pre-populated with every figure traced.
The whole flow, in one chart: what gets read, what your team builds, what lands on your desk, and how it gets sharper every month. We build the first version with a few firms, free during the beta, around workflows they actually need.
The repetitive grind (spreading, covenant checks, report assembly) runs on schedule. Your team reviews flagged exceptions instead of rebuilding spreadsheets.
Every number is recomputed, never copied: a changed EBITDA definition, a missing certificate, a stale mark gets caught and flagged. And nothing is overwritten: if a borrower restates EBITDA, you still see the original, when it changed, and which memos relied on it.
Checked against your closed quarters before go-live, and every correction is remembered for your firm. Non-recurring questions become agents your team builds, on the same machine, with the same sourced output.
Your definitions, your covenant language, your report formats, your teams. Your team builds its instance on your data, and the system shows the evidence behind every deduction before anything ships.
Private lending, private equity, venture, family office. The ontology is defined per firm, not per strategy, so a borrower’s covenant and a portfolio company’s ownership run on the same machinery, and no engine limits your metrics.
Your provider’s API under your own contract, an endpoint in your own tenant, open weights on your own hardware, or a model your team built. We never hold keys, we never mark up tokens, and nothing is sent to a model your compliance team didn’t approve. Every figure is identical with the generative layer switched off entirely.
Every ratio is computed by deterministic code and clicks through to its source clause. Nothing is overwritten, so a restated figure keeps its original and everything that relied on it. It is the trail your auditors, and your LPs, wish every manager had.
We won't pretend the agents are perfect. Here's the real promise: they read, cross-reference, and compute across your whole book in minutes, show you where every number came from, compute with exact code, and flag clearly what they couldn't confirm. Your team decides, with the evidence in front of them.
Built for the way a fund's IT, compliance, and LP due diligence teams actually evaluate a vendor.
The model has no database credentials. Agents read through a permissioned gateway and can’t change, move, or delete anything.
Single tenant, in your own subscription, behind your private link. Your documents never leave your perimeter.
Your own AI subscription, an endpoint in your tenant, or a model your team runs. We never hold keys, and nothing goes to a model you didn’t authorize.
Your data and documents are never used to train any model, ours or anyone else's.
Encrypted in transit and at rest, isolated per firm, with scoped access you can revoke at any time.
A full trail of what each agent did, and every figure traces back to its source page.
Penetration testing and SOC 2 are on the near-term roadmap, before scale. Before any data touches the system, we sign an NDA and walk your security and IT teams through the architecture.
A98 is built by a credit and data professional who underwrites and monitors lower-middle-market deals every day, and writes the code behind it. The founding beta is deliberately small: five funds whose workflows shape what the machine becomes.
For GPs, CFOs, COOs and the analysts who do the work. Tell us which workflow you’d rebuild with your own rules and approvals, and we’ll show you what it would look like on your own numbers.