Consolidation
142 positions, 118 funds
The 142 positions across the book resolve to 118 unique funds; some are held through multiple entities, and the store ties those rows back to one fund automatically.
One governed store over the whole book, documents that read themselves, and answers that cite their source. Every capability below earns its place by making a real question answerable, and keeps a human in the loop wherever judgment belongs.
Private equity and commercial real estate share a single store, with fund-level consolidation from day one. The store keeps time-series, not snapshots, so both "the portfolio as of any date" and "what changed this quarter" are answerable; today they are not, because NAVs blend quarter-ends.
The 142 positions across the book resolve to 118 unique funds; some are held through multiple entities, and the store ties those rows back to one fund automatically.
Full look-through exposure across the whole book comes back in a single query, PE and CRE together, rather than being reassembled by hand across sheets.
Because the store records figures over time rather than overwriting them, "the portfolio as of March 31" and "what moved this quarter" are both real questions with real answers.
A proven taxonomy of about a dozen document types (GP and Goldman statements, capital calls, distributions, and more) has been validated against real mail. Extraction is engineered to be trustworthy before it is fast: senders are gated, confidence is earned across runs, and failures route to a person rather than guessing.
A sanctioned-sender whitelist admits documents for parsing. Non-whitelisted senders are rejected outright, never parsed, so nothing off-list reaches the extraction path.
Confidence is multi-signal, based on agreement across multiple runs rather than a single model's self-reported score, so a lucky guess cannot pose as certainty.
Schema-validation failures route to review with a best-effort extraction attached, so a person picks up exactly where the machine stopped.
Content-hash dedupe means the same document arriving from several inboxes is absorbed a single time, without double-counting a call or a distribution.
Every extraction agent ships through a shadow-then-canary rollout, proving itself against live mail before it is trusted with the book.
The document taxonomy covers roughly a dozen types validated against real mail, so the common shapes of the firm's inbox are handled, not hypothesized.
Extracted figures surface in a Slack review queue as claim-and-resolve cards. This human layer is the verification step before anything enters the store; the machine proposes, a person disposes.
Every extracted figure appears in #ingestion-review as a claim to confirm, correct, or reject. A person is the last step before a figure is real, so verification is a habit, not an afterthought.
A weekly random-sample audit, on the order of 5 percent, tests the automated scorer itself. The system that judges confidence is held to the same scrutiny as the documents it reads.
Each figure carries its source document, data class, and as-of date: figure-to-document links, not cosmetic notes. The store reconciles to the systems of record, complementing them and never replacing them, and refuses to migrate quietly if the numbers do not agree.
Every figure links to the document it came from, along with its data class and as-of date. These are load-bearing links back to evidence, not decoration.
The store reconciles to the accounting system of record (Sage Intacct) and the investor portal (Juniper Square), complementing both. It is a governed view over the book, not a replacement for either.
A set of reconciliation checks gates the migration off the legacy spreadsheet and fails loudly if counts and totals do not tie out, so the switch happens only when the numbers agree.
Natural-language query from Slack or a dashboard, with answers that come back cited by source and as-of date. Clearance is enforced in the answer itself: an above-clearance ask returns a dashboard deep link, not the figure.
Ask in #portfolio or on a dashboard in plain English. No query language, no spreadsheet spelunking; just the question you would ask a colleague.
Every answer comes back with its source and as-of date attached, so a number is never separated from the evidence and the moment it was true.
When an ask sits above your clearance, the platform returns a dashboard deep link rather than the number. Refusal is a governed answer, not a dead end.
The live pipeline runs in Linear across two teams, an Alternatives track and a CRE track, one issue per deal. Agents do the busywork under a dedicated identity, but only a human changes a deal's stage, and no book-of-record position exists until a person has checked it.
The pipeline lives in Linear across an Alternatives track and a CRE track, with one issue per deal. Agents create, comment, label, and attach under a dedicated identity, so their work is visible and attributable.
Agents never advance a deal on their own; only a person changes a deal's stage. Deal and deal-stage-history records live in the store, so the pipeline is queryable, not just visible.
The morning brief gains a live pipeline diff, and capital-weighted pipeline, conversion, time-in-stage, and kill analysis all come from the store with citations.
Closing a deal spawns a human-owned handoff checklist. No book-of-record position row exists until a human has checked it, so the record of truth is never created by a bot.
Process channels are durable and single-purpose; per-deal channels are ephemeral, created at screening and archived on decision. Threads act as working sessions, corrections are first-class, and posts are batched by default so only what is urgent interrupts.
Ask in plain English and get an answer with its source and as-of date attached.
Extracted figures wait here as claim-and-resolve cards for a person to confirm.
Extracted capital calls and distributions post with their source documents linked.
A weekly report on what arrived, what has gone stale, and what is missing.
Coverage and loan-to-value thresholds and upcoming maturities, surfaced early.
Ephemeral channels open at screening and archive on decision, so the surface stays clean.
A thread is a working session: an artifact is posted and iterated in-thread with version notes, so the history of a decision stays with the decision.
When a source finalizes late, the bot edits the original post and threads a correction notice, so a stale figure is never left standing silently. Posts are batched by default; only deadline-class items post on their own.
The economics of the book are computed by a deterministic, versioned engine, never by a model. Every figure below is reproducible from the deal's terms, so the numbers are auditable rather than approximate.
Per-investor waterfall and preferred-return tracking, computed from each deal's terms rather than maintained in a shadow spreadsheet.
Bank and cash balances alongside expected flows, so the live cash picture and what is coming are in one place.
GP fee and promote economics computed the same way every time, by the deterministic engine and never by a model.
A side-letter register captures non-standard terms, so the exceptions that today live in a few people's heads become findable with their record.
The connectors to the systems of record are built a single time and shared, each access-gated by role with one credential store. Where a source offers no API, the plan is export parsing rather than blocking on it.
The accounting AI Gateway reads Sage Intacct read-only, so the store reconciles against the ledger without ever writing to it.
The investor-portal connector reads Juniper Square, keeping the store aligned with what investors see.
Where a source has no API, such as Goldman, the plan is export parsing rather than blocking. A missing API slows a source down; it does not stop the book.
The same foundation reaches further without changing shape: consolidation across the firm, alerts that arrive inside the briefing rather than in yet another notifier, and tools to model what-ifs.
Firm-wide consolidation extends the single look-through view beyond the initial book, on the same governed store.
Alerts flow into briefings rather than a separate notifier, so what matters arrives where you already look instead of a new place to check.
Scenario tools let you model alternatives against the governed store, with the same provenance and clearance rules in force.
Adoption is staged and safe by construction: the data layer first, then ingestion with a human in the loop, then query and dashboards, then the extras. Each stage earns the next, and every access follows the rules the firm already ratified.