From pilot to production — what the work looks like.

Challenge

Finance keyed vendor invoices by hand, reconciled them against purchase orders and contracts, then manually produced customer invoices — slow, error-prone, and hard to audit at month-end.

Approach

An agentic intake pipeline extracts line items from vendor invoices, matches them to POs and contract terms, and drafts the corresponding customer invoices — with the AI proposing values, never finalizing them.

Controls that made it safe

• Deterministic validation (format, PO existence, totals, contract rules) before any posting
• Provenance on every extracted field — source document and confidence
• Human review for high-value or low-confidence items
• Immutable audit trail of every posted entry

Outcome

Challenge

A core operational workflow spanned several systems and required staff to copy data between them, chase approvals, and handle exceptions by hand — capacity-limited and inconsistent.

Approach

I mapped the end-to-end workflow and introduced agents to handle the routine path — reading inputs, updating records, routing approvals — while exceptions escalate to a human queue.

Controls that made it safe

• Risk-tiered automation: routine auto-handled, consequential steps gated
• Every action scoped to verified identity and least privilege
• Reversibility — any step can be paused and rolled back
• Full decision log for traceability

Outcome

Challenge

Onboarding a new customer or service meant coordinating setup across CRM, billing, and operational systems — days of manual steps, with frequent rework when something was misconfigured.

Approach

I orchestrated provisioning as a guided, mostly-automated flow: agents assemble the configuration, validate it against business rules, and activate across systems once every check passes.

Controls that made it safe

• Validation gate before activation — no half-provisioned states
• Tenant-scoped data access throughout
• Human approval for non-standard configurations
• Audit trail from request to activation

Outcome

Challenge

Specialists spent hours reading dense regulatory, and operational documents to determine what applied and what action to take — a bottleneck where mistakes carry real compliance risk.

Approach

An LLM-assisted layer interprets the source text and proposes the applicable rules and required actions, surfaced to specialists with citations — accelerating interpretation without removing human judgment.

Controls that made it safe

• Provenance over confidence — every interpretation cites its source passage
• Deterministic checks against authoritative reference data
• Human-in-the-loop for any consequential determination
• Immutable record of what was interpreted and decided

Outcome

Challenge

A legacy estate — monolithic, hard to change, expensive to run — needed to become cloud-native and AI-ready, but a multi-year rewrite is risky and slow with traditional methods.

Approach

A pragmatic, phased migration to cloud-native microservices, accelerated with AI-assisted development and quality gates — keeping the legacy system stable in a transition-state architecture while new modules ship.

Controls that made it safe

• Quality gates (tests, scans) on every AI-assisted change
• Strangler-fig transition — legacy stays stable while functionality moves
• Observability and rollback at each cutover
• Measurable coverage and defect tracking

Outcome