Alternative asset fund operations · applied AI for fund accounting
CPA (NY, inactive) · MAcc · FMVA · CAIA Level 2 Candidate
I spend my working hours as a lead technical reviewer for NAV production across hedge fund, private equity, credit, and crypto strategies. I spend my own hours building the tooling that job keeps proving should exist — a suite of fund-operations tools with a single architectural rule: deterministic code computes; language models only extract, explain, and disagree.
Seven tools, built end to end. Every financial figure is recomputed by deterministic code before any model reasoning touches it.
| Tool | What it does | Status |
|---|---|---|
| FundLens Audit | Four role-specialized agents (LP counsel, fund accountant, regulatory reviewer, MFN auditor) audit a complete fund document set; a synthesizer surfaces inter-agent disagreement as a first-class finding | Deployed |
| FundLens Capital | LP capital lifecycle and waterfall modeler — call pacing, preferred accrual, peak funding need, J-curve. Engine v1.2.0 | Deployed |
| DraftLens | Compares prior- and current-year draft financial statements; flags variances against dual thresholds and maps findings to ASC 946, 820, 825, 326, 275 | Deployed |
| FundLens | Document intelligence for LPAs, PPMs, capital account statements, and side letters via PDF vision, with schema-validated extraction | Deployed |
| FundLens Forecast | Scenario projection across IRR and TVPI trajectories under varying assumption sets | Deployed |
| FundLens Waterfall (DesoFall) | LPA-driven distribution calculator; its tier engine was extracted into the shared library the suite now runs on | Deployed |
| LedgerBridge | Deterministic field-mapping engine normalizing fund accounting exports into a canonical schema across 11 entity types, conforming to AICPA Audit Data Standards | Stage 1 |
The interesting problem in this domain isn't generating output — it's proving the output is correct when errors are silent, plausible, and compounding.
- Deterministic core, agentic judgment. All cash-flow math and scenario arithmetic live in pure, unit-tested code. Every parameter is user-visible and defaulted from documented sources. Models never select a parameter.
- Design red-teaming before the build. Attacks on the design get written down and dispositioned — shipped, deferred behind a gate, or refused — before code exists.
- Seeded-discrepancy testing. Known defects injected into fixtures, with a single detecting assertion per case.
- Cross-model adversarial validation. A second frontier model with no repository access recomputes the full waterfall from dated primary events and files quantified findings. On FundLens Capital this surfaced two preferred-return convention defects and moved roughly $373K–$424K of carry from the GP back to the LP on the test fixture.
- Documented limitations. Known exclusions — recycling, subscription credit lines, NAV facilities — are named in scope docs with reserved input flags rather than left as silent assumptions. In this domain a documented limitation is credibility; a discovered one is disqualifying.
TypeScript · Next.js · Python · SQL · Anthropic SDK · Neon/Postgres · pytest · Vitest · Vercel
~7 years across fund administration, private equity fund accounting, and financial statement audit. Authored the GCES-V prompt engineering and GCES-B agent instruction frameworks, adopted as business unit-wide standards for financial operations AI interactions.
Everything in this profile was built on personal time and personal equipment, using synthetic fund data throughout. No employer or client data is present in any repository here.