Every entry,
under observation.
An AI observatory for sovereign finance. Detects anomalies, matches reconciliations, extracts bilingual invoices — with full transparency and an immutable audit trail.
Eighteen agents, one ledger of trust
Each built on LangGraph, sharing one data layer and one audit log — and every one explains itself.
Detect & investigate
Anomaly Detection
Duplicate payments, after-hours postings, and statistical outliers — each with a calibrated, auditable narrative.
Benford Forensics
First-digit distribution analysis on vendor spend to surface fabricated or manipulated amounts.
Vendor Risk Scoring
Ranks vendors 1–10 on payment history, late-pay rate, and anomaly hits, with the evidence behind every score.
Reconcile & process
Bank Reconciliation
LLM-driven fuzzy matching of statement lines to open invoices — Arabic ↔ English names, partial payments.
AP Invoice Intake
Drop a PDF, get a draft vendor bill in Odoo: vision extraction, vendor match, GL coding, confidence per field.
3-Way Match
Invoice ↔ purchase order ↔ goods receipt. Flags quantity, amount, and vendor drift before payment.
Forecast & collect
Cash-Flow Forecast
Projects the net cash position across 30 / 60 / 90-day windows from live AR, AP, and bank float — narrated.
AR Collections
Ages overdue receivables and drafts a bilingual reminder per invoice, escalating by severity.
Procure to pay
P2P Control Tower
The whole funnel from request to payment in one view, with a health score that names the components it cannot measure.
Sourcing Agent
Ranks who to invite on demonstrated delivery history and scorecard, and says plainly when no competing quote is on file.
Negotiation Agent
Targets a price the supplier already charged, draws the discount authority floor, and escalates rather than exceeding it.
Delivery Intelligence
Predicts a slip only where the supplier has enough completed deliveries behind it, and reports its own coverage.
Payment Control
Six checks before every payment. A check that could not run is shown as such, never as a pass.
Supplier Intelligence
A scorecard per supplier from real spend and billing history, with unmeasurable components named rather than scored zero.
Assist & query
Journal-Entry Assistant
Describe a transaction in plain language; it proposes a balanced, GL-coded journal entry for review.
Text-to-SQL
Ask a finance question in English or Arabic; it returns a read-only, sandboxed query and the answer.
Copilot Orchestrator
A ⌘K copilot that composes answers across every agent and logs the conversation to the audit trail.
Fixed-Asset Register
Tracks the asset register and depreciation schedule, flagging gaps and anomalies in the books.
From request to payment, every decision observed
Twelve stages above Odoo. At each one the platform states what it is permitted to do, and most of the time that is observe or recommend.
- 01
Request
Reads the requirement, checks it against open orders and stock before anything is bought.
Recommends - 02
Approval
Routes on amount, category and supplier risk, and tells the approver why they are being asked.
Human approval required - 03
Sourcing
Ranks suppliers on delivery history and scorecard, and flags a single-source requirement.
Recommends - 04
RFQ
Drafts the invitation and the supplier shortlist. Issuing it to Odoo is a separate, named step.
Prepares a draft - 05
Negotiation
Sets a target from prices the supplier already charged, inside a declared discount authority.
Prepares a draft - 06
Purchase order
Validates supplier, price, budget and policy. Never confirms a final order on its own.
Human approval required - 07
Delivery
Predicts a slip per supplier, and says when its history is too thin to predict one.
Observes - 08
Goods receipt
Compares what arrived against what was ordered, and against how this supplier usually performs.
Observes - 09
Invoice
Extracts every field with a confidence score; anything under threshold goes to a reviewer.
Recommends - 10
3-way match
Invoice against order against receipt. Green, amber or red, with the figures behind the verdict.
Recommends - 11
Payment control
Six checks before release, including duplicates and supplier bank changes. Never pays autonomously.
Human approval required - 12
Optimisation
Tracks price movement, off-contract spend and concentration, and quantifies only what the data supports.
Observes
What this adds over a dashboard
A payment gate that admits what it does not know
Every check returns pass, fail or not checked. A blocking check that could not run is called out, because releasing on an unrun check is a decision someone should make deliberately.
Policy and autonomy, enforced before the write
An autonomy level per agent and policy rules editable without a deploy. The stricter of the two wins, and the gate sits in the write path, not on a dashboard.
Savings quantified only where evidence exists
Price movement is measured supplier by supplier over time. Realised and negotiated savings report as unmeasurable until a baseline is tracked, rather than as zero.
Provenance on every figure
Live means counted from Odoo on this request. Agent means replayed from a decision already in the audit log. The badge is on the number, not in a footnote.
AI that never asks you to just trust it
For a government audience, trust is the product. Every decision exposes its reasoning, is logged, and waits for a human where it matters.
- Append-only audit trailEvery agent decision and copilot answer is written to an immutable log — the auditor-general’s single source of truth.
- No black boxEach finding carries a “Why this decision?” panel: the reasoning, a calibrated confidence, and the source records it cites.
- Human-in-the-loopHigh-stakes actions pause for human approval via graph interrupts. The AI proposes; a person disposes.
Matched on vendor_id + amount within the 7-day duplicate window. Reference similarity 0.91. No reversing entry found. Routed to human approval before any hold.
Observe. Explain. Approve.
Mirsad sits on top of your existing Odoo 18 & 19. No data migration, no plugins.
- 01ConnectMirsad talks to Odoo 18 & 19 over JSON-RPC. No Odoo plugins, no data migration, read-only by default.
- 02ObserveLangGraph agents scan journal entries, vendor bills, and bank statements continuously. Every rule is explicit.
- 03ExplainEvery finding ships with evidence, a calibrated confidence, and source references — in Arabic and English.
- 04ApproveHigh-stakes actions pause for human approval via graph interrupts. No black-box autonomy.
Built for the sovereign stack
Government buys trust and sovereignty, not “AI features.” Mirsad is engineered for both.
Sovereign cloud or on-prem
Runs in your region — AWS Bedrock, Azure Sovereign, or fully air-gapped on a vLLM you control. Data never leaves your tenant.
No model lock-in
A single swap point switches the LLM between Anthropic, Bedrock, Azure, or in-region Llama / Qwen. Agent logic never changes.
Read-only by default
Mirsad observes. Any write — a draft bill, a reconciliation note — is proposed, logged, and gated behind a human.
Always auditable
Append-only audit log with per-decision reasoning, confidence, and source records. Designed for an INSERT-only role and hash-chain.
Bilingual & RTL-native
Every screen flips to full right-to-left. Findings are narrated in Arabic and English — Hijri-aware where it matters.
No data migration
Plugs into your existing Odoo 18 or 19 over JSON-RPC. Nothing to install inside Odoo; nothing to move out of it.
Questions an auditor-general asks
The answers finance and audit teams need before any commitment.
Where does our data live?
Can it run without a foreign LLM or the internet?
Odoo 19 ships its own AI. Why Mirsad?
Does it modify our Odoo books?
Is every AI decision auditable?
What about Arabic?
How long does it take to deploy?
See Mirsad on live Odoo data.
Pre-seeded vendors, planted anomaly patterns, agents running live. No production access required.