AI financial systems | Agentic infrastructure | Quantitative research · NYU Stern MBA · previously founded and exited an enterprise compliance business

I build and evaluate financial systems where software can act on money, data, or infrastructure, and where authority, evidence, and financial semantics matter as much as model capability.

My background spans an MBA from NYU Stern, structured credit, financial-crime investigations, private-equity underwriting, enterprise AI, and independent technical research. I previously founded and exited an enterprise compliance business, including an IP sale and a separate sale of the operating company.

Portfolio status: Independent systems research and prototypes; some concepts were briefly tested as products, but none became an ongoing commercial operation. Implementation repositories are private; these public case studies document architecture, decisions, corrections, limitations, and validation evidence.

Authorization, approval, capability isolation, economic controls, reconciliation, and auditability for AI agents

Designed a controlled-execution architecture for software agents operating across consequential actions. The system combines MCP, OAuth 2.0 and PKCE, scoped permissions, acquisition entitlements, spending controls, digest-bound human approval, idempotent state transitions, signed runtime distribution, audit evidence, and a constrained execution runtime.

Evidence: 161 registered mutation/recovery cases, automated source-mutation testing, signed runtime-distribution checks, and one reviewed hosted execution path.

Focus: agent authorization, secure execution, transaction controls, human-in-the-loop decisions, PostgreSQL, TypeScript, Next.js.

Temporal financial graphs and probabilistic cash-flow forecasting over linked-account systems

Built a financial-graph representation of accounts, transfers, repayments, recurring behavior, and cash flow. The forecasting architecture separates deterministic known obligations from learned residual uncertainty and preserves a clear boundary between observed facts, model outputs, and real-world reliability claims.

Evidence: 700 synthetic users, 439,092 generated transactions, 24 training epochs / 13,128 steps, and point plus q10/q50/q90 forecasts across 90, 180, 365, and 720-day horizons. Evidence remains offline rather than production-calibrated.

Focus: financial graphs, transaction reasoning, probabilistic forecasting, Plaid, PostgreSQL, TypeScript, Python model workers.

Scientific quantitative research infrastructure and adversarial model evaluation

Built a short-horizon research laboratory designed to discover and falsify market effects under point-in-time information boundaries. It combines immutable evidence, canonical PostgreSQL memory, typed OCaml research, hostile methodological testing, deterministic scientific memory, and explicit evidence-exposure rules.

The programme has progressed into prospectively registered statistical-methodology and publication work while preserving a strict boundary between research evidence, trading hypotheses, and production authority.

Evidence: 200 liquid U.S. equities, approximately 720 calendar days of history, 7,659,672 frozen baseline bars, more than one thousand automated scientific and systems checks, and a current registered fixed-k DCDR/ECC studentizer paper at Baseline V1.

Focus: OCaml, PostgreSQL, intraday market data, experimental design, econometrics, scientific memory, numerical research, reproducibility.

Consent, provenance, canonical financial semantics, and institution-neutral interfaces

Designed a provider-neutral financial-data architecture with REST and MCP interfaces over one application core. It emphasizes exact money representation, provenance, explicit absence semantics, durable consent state, idempotency, auditable authorization, deterministic financial intelligence, and replaceable institution adapters.

Evidence: locally runnable PostgreSQL-backed sandbox, 12 canonical capabilities and authorization scopes, six MCP resource templates, cross-interface consent enforcement, durable idempotency/audit, and institution API conformance tooling.

Focus: financial data contracts, consent, provenance, MCP, REST, PostgreSQL, TypeScript.

• Financial semantics before convenience. A technically valid number is still wrong if its economic meaning is wrong.
• Authority must be explicit. Reading, spending, mutating, approving, and deleting are different powers and should be modeled that way.
• Evidence classes should remain distinct. Synthetic, sandbox, offline, live, and production evidence support different claims.
• Corrections are part of the record. A result that changes after stronger testing should be narrowed or rejected, not cosmetically rescued.

These repositories are public case studies of independent systems work. Some concepts were briefly tested as products, but none became an ongoing commercial operation. The underlying implementation and active research repositories remain private.

The public case studies expose architecture, decisions, validation evidence, corrections, and current system boundaries without publishing proprietary code, customer data, credentials, or security-sensitive operational detail.