What's real

Priya Raman · Quant & DX Champion P

Leadership · The skeptic's page

This is not a mock. Here is the proof.

A low-code AI platform is easy to demo and hard to trust. So this page does not describe the controls — it runs them, right now, against the live database, and shows you the result. The cryptography below is re-verified on this request; the constraints and triggers are read out of the database's own catalog; the denials and halts are real rows. Where something is a stand-in, it says so plainly.

Cryptographic sign-off

LIVE · re-verified now
VALID HMAC re-derives over the signed graph hash

TECH_STANDARDS gate on Cross-Sell Intelligence Hub, signed by Raj Patel.

Tamper test: a flipped field re-verifies FALSE

Four-eyes, in the database

LIVE · schema catalog
ENFORCED

A trade cannot be approved by its own trader — enforced below the application, so no bug or script can self-approve a booking.

`approver_user_id` is null or `approver_user_id` <> `trader_user_id`

26 bookings actually approved by a second human.

Append-only sign-off ledger

LIVE · schema catalog
2 triggers

A signed approval you can later edit is worthless. BEFORE UPDATE and BEFORE DELETE triggers make the ledger immutable by law of the database, not by convention.

trg_mos_signoffs_no_delete
trg_mos_signoffs_no_update

The AI never writes blind

LIVE · real gate firings
4 write attempts DENIED, of 35 calls

Every call to Murex / TLM / Bloomberg / Progmat / ERP passes the MCP broker: only allow-listed operations run, and a write against a read-only connector is refused — and the refusal is audited, because "we declined the dangerous thing" is exactly the evidence a regulator wants.

Runs halt for a human

LIVE · run ledger
20 runs on the ledger

3 are parked PENDING_HITL — halted at a human gate, state persisted, waiting on a decision. The engine never blocks a thread; it stops and resumes.

High-Risk Logic Auditor

COMPUTED · live scan
0 blind spots across 15 PROD apps

Scans every live app graph for an irreversible auto-executing action — an on-chain settlement, a tokenisation, an MCP write — with no human gate or monitor anywhere in the graph. Not one PROD app auto-executes uncontrolled.

Capability matrix — what is real, honestly

Every material capability, tagged by how real it is today

5 LIVE 3 COMPUTED 3 SIMULATED
CapabilityStatusEnforced byWhat that means
Cryptographic governance sign-offs LIVE HMAC + append-only DB triggers Every sign-off is HMAC-bound to the exact graph hash and re-verified at promotion; the ledger is append-only by SQL trigger.
DAG execution with human-in-the-loop LIVE Workflow engine + run ledger Runs walk a topological order and halt into PENDING_HITL at a human gate, persisting state — never blocking a thread.
MCP allow-list & write denial LIVE McpBroker + invocation audit Only allow-listed operations run; a write against a read-only connector is DENIED and the refusal is audited.
Four-eyes on trade approval LIVE Database CHECK constraint mos_chk_trade_four_eyes forbids a trade whose approver is its own trader — enforced below the application.
Immutable audit trail LIVE Append-only audit log Every governed action writes an audit row; overrides require a justification enforced by a DB CHECK.
Risk, pricing & VaR calculation COMPUTED Domain services over demo data Real algorithms (VaR, DV01, Black-Scholes, stress shocks) computed over the seeded book — deterministic, not market-connected.
Cross-sell whitespace intelligence COMPUTED Ontology reasoning Genuinely reasons over the entity graph to find loan clients with unhedged FX — on the demo loan book.
Process mining & efficiency proposals COMPUTED Node-result aggregation Aggregates the real run journal to surface bottlenecks and automatable gates.
Murex / TLM / Bloomberg / Progmat / ERP SIMULATED Deterministic connectors Connector responses are faithful, deterministic stand-ins. The MCP boundary, allow-list and audit around them are real.
Market & reference data SIMULATED Seeded curves & quotes Curves, quotes and FX are seeded fixtures, not a live feed.
The AI copilot / LLM reasoning SIMULATED Scripted assistant The copilot is a scripted stand-in; a real deployment would place a governed LLM behind the same MCP boundary.

What a real deployment still needs

The honest gap between this proof-of-concept and production

Real connectivity

The MCP boundary, allow-list and audit are real; the connectors behind them are deterministic stand-ins. Production swaps each for the live Murex / TLM / Bloomberg / Progmat / ERP endpoint — the safety layer does not change.

A governed LLM

The copilot is scripted here. A real deployment places a genuine LLM behind the same MCP boundary and the same governance pipeline, so the model can read but never act un-audited or un-approved.

Production security

The signing key is an in-code stand-in for an HSM/KMS-held key; auth is a persona switcher. Production wires SSO, real RBAC and a hardware-held key — a confined change, since the crypto already routes through one service.