Skip to main content
§ Product

Conversational Portfolio Assistant

An LLM-powered natural-language interface on top of the Multi-Entity Consolidation Platform — accepts questions in plain English, queries the underlying data, returns answers with the supporting numbers and audit links.

Engagement
4–8 week build · ongoing model maintenance
Built for
Principals · CFOs · Operations staff
§ Problem

The consolidation platform has the data, but accessing it still requires knowing how to use the dashboards, or asking the CFO. Principals don't want a dashboard; they want to ask a question and get an answer.

What this is

The principal-facing layer of the suite. Three components:

  • Natural-language query. Question → SQL → answer, with the LLM as the translation layer. Per-FO fine-tuning on the question patterns that actually come up.
  • Permission-aware execution. All queries run in the asker's permission context; the LLM doesn't see data the user can't see.
  • Auditable answers. Every answer includes the supporting numbers and links to the source transactions. Trust is built through transparency, not through marketing.

How it's built

LLM layer (Claude-class commercial API, or Llama-class on-prem fine-tune) with a query-construction layer that translates natural language to SQL against the canonical schema. Permission enforcement at the database layer (RLS), not the LLM layer — security guarantees don't depend on model behavior. Calibration tracked against a per-FO question set.

What you get

  • The assistant deployed to your FO — Slack, Teams, or web UI per preference.
  • Per-FO fine-tuning on your team's question patterns.
  • Auditable query path from question to answer.
  • Calibration tracking and drift detection.
  • Ongoing fine-tuning as new question patterns emerge.
§ How we engage

Engagement is shape, not list.

Length and price are functions of the data and the destination. The shape below is the typical engagement.

Length
4–8 week build · ongoing model maintenance

Scoped during the discovery call against the actual data and the operation it integrates with.

Lead
Bogdan

Principal engineer. Architecture and most code ships through one keyboard.

Cadence
Async, weekly

Written updates between, calls when the decision needs the room.

Bar
Production

Async correctness, capacity under burst, observability at every boundary.

§ Questions

What buyers ask about this one.

  • How does it handle the privacy partitions?

    Every query runs in the user's permission context. A principal asking 'what's our cash position?' sees their own family's view; the same query from the CFO sees the consolidated picture. The LLM doesn't bypass the access controls — it queries against the row-level-security-filtered data layer.

  • What if the LLM gets the answer wrong?

    Every answer comes with the underlying query and the data sources cited. The principal can audit the path from question to answer. Where the question is ambiguous, the assistant asks for clarification rather than guessing. And we measure accuracy on a held-out question set per FO — model drift is detected before users notice.

  • Which LLM do you use?

    Commercial-API models (Claude-class, GPT-class) where the FO permits, fine-tuned on the FO's question-and-answer patterns. Where the FO requires on-prem inference, we deploy a Llama-class model with the same fine-tuning approach. Either path is documented per engagement.

  • What questions can it actually answer?

    Anything answerable from the consolidation platform's data — cash positions, capital-call schedules, distribution histories, allocation by asset class, exposure to specific managers or sectors, year-over-year comparisons. Where the question requires forecasting or scenario analysis, the assistant routes to the Liquidity & Cash Forecasting Engine.

  • Pricing?

    Scoped against question volume and the on-prem-vs-cloud decision. Discovery call covers both.

§ The next step

If the deliverable matches the gap, the next step is one call.

We'll scope length and price against your data and the operation it integrates with. No retainer, no fishing.

Bogdan and team · async-first · OP—2026