Skip to content

Proof

What we have built, and how we know it works.

Client engagements are described by domain, not by name — the work is real and the confidentiality is as well. Founder history is attributed to the firm that held the contracts.

Redacted architecture of a reporting platformMultiple MySQL RDS databases, an on-premise database and partner APIs are replicated by AWS DMS into a Redshift warehouse. Airflow on MWAA orchestrates the loads in Python. Sigma reads the warehouse for dashboards, and a document path generates invoices, receipts and annual 1099s per client from configuration.SourcesReplicationWarehouseConsumersMySQL RDSseveral accountsOn-premise MySQLoperational systemsPartner APIsthird-party feedsAWS DMSCDC replicationRedshiftsingle source of truthSigmadashboardsDocumentsinvoices, receipts, 1099s~25 clientsper-client configurationAirflow · MWAAPython DAGs schedule and verify each loadRemoved before publication· client and provider names· hostnames and account ids· volumes and ratesBatch paths· annual 1099 run for every service provider· monthly invoices reconciled against receipts
The reporting platform as built, simplified. Structure and technology are the delivered ones; anything that would identify the client has been removed rather than the diagram withheld.

Report excerpt

What finance signs off

PartnerInvoicedReceiptedVariance
Partner A██████$0.00
Partner B██████$0.00
Partner C██████$0.00
Partner D██████$0.00

Figures redacted. The columns are the ones delivered, and the variance column is the point: a partner whose receipts do not match its invoice is a finance problem, not an IT one.

Workstreams on one engagement, and one proof of concept

The same client kept coming back, which is the only reference that matters.

  1. 01

    Data lake proof of concept

    Ingest from RDS endpoints, on-premise MySQL and third-party APIs into one source of truth for reporting.

  2. 02

    Support emotion analysis

    Tickets pulled from Redshift, personal data redacted before any model sees it, triaged by complexity across three model tiers, analysed for emotion, and delivered as a ten-tab dashboard covering cost, compliance and model performance.

  3. 03

    Cloud infrastructure

    EC2 fleets hosting the document generation and reporting services, containerised with Docker throughout.

  4. 04

    Documents and reporting

    Invoices, receipts and annual 1099s for roughly 1,200 service providers across 25 clients, generated from per-client configuration.

  5. 05

    Visualisation platform

    A component-based BI framework for a second client, in a different sector, that non-programmers could author.

  6. 06

    Governed FinOps assistant

    A proof of concept in which plain-language cost questions are answered by calling governed tools against a data catalogue — with authentication and scoped SQL — rather than by letting a model invent a query.

Live

National roadside assistance · production

A multi-terabyte data platform behind a national roadside network

The problem

Claims, dispatch, billing and partner reporting had grown into hundreds of scheduled shell scripts with no owner, no lineage and no failure alerting.

What we built

Change-data capture into Redshift, roughly 110 orchestrated pipelines, a semantic layer over the warehouse, and partner reporting delivered as files to a schedule.

Why it matters

Revenue recognition, provider invoicing and insurance receipts became traceable — and the legacy cron estate could be retired deliberately rather than all at once.

Pre-production

Secure ingestion of sensitive financial documents

OctoLedger: statements in CSV, OFX and PDF are quarantined, malware-scanned, parsed with bounded limits and imported atomically. Tenant isolation is enforced across 34 tables, with isolation checked by tests that attempt against a four-hour objective.

Not launched. Presented because the ingestion pattern is exactly the one claims work needs.

In development

Citations that do not resolve are rejected

Navelyra: generated text is bound to facts a practitioner selected, and a citation that does not resolve is rejected before it is persisted. Prohibited claims are caught in code. A human approves the exact version that ships, and that approval is recorded against a content hash.

A measurement, with its context

What a restore actually took, not what it was supposed to take.

14m 16s

Measured restore

The production-shaped database restored into a disposable instance, with row counts verified against source

4h 00m

The objective

PostgreSQL 18, encrypted at rest. Measured during pre-production hardening, against production-shaped data

2026

When it was run

Recorded with the restore log, not inferred from configuration

Founder history · M-Cubed Information Systems, 1985–2015

Thirty years of systems that had to reconcile.

These were M-Cubed’s contracts, not OcularIT’s. They are here because they explain where the discipline comes from.

U.S. Department of the Treasury

GTAS financial reporting — prime contractor

Trial-balance data collected from federal program agencies, validated against central-accounting data and Standard General Ledger rules, and used in preparing the Financial Report of the U.S. Government. Performance-based contract with earned-value reporting.

U.S. Census Bureau

DADS-II — largest subcontractor to IBM

American FactFinder: guided search and navigation across more than twenty thousand statistical tables and hundreds of thousands of geographic divisions, on tabulation infrastructure built for the decennial census.

U.S. Army

Victim and Witness Information Portal

A case-management portal delivered as SaaS on AWS GovCloud — before GovCloud was a common choice — with Section 508 accessibility obligations.

Process and accessibility

CMMI Level III appraised; Section 508 delivery

Delivered under a CMMI Level III appraisal at the time, with earned-value reporting under a quality assurance surveillance plan, and 508 accessibility verified on a public-facing government site.

CMMI is an appraisal held at a point in time. M-Cubed’s appraisal expired with the company, which was sold in 2015; OcularIT does not hold a current CMMI appraisal and does not claim one.

The fastest way to judge any of this is to ask us a hard question.

Ask one