Lending Product Leader · AI-native across software & operations

Deep lending expertise — now AI-native, across software and operations.

I'm Nick Allen. 20+ years in Australian lending — credit, operations, securitisation, product. I designed Data Sense, the Document Intelligence rule engine now running in production. Through 2026 I've gone AI-native — applying it to how lending operations run and how software ships, with a stack of AI subagents and MCP integrations that run the full lifecycle, from refinement|

11
Agents — 7 run standalone
Data Sense
Designed · live in production
~45
Rule templates shipped
20+
Years lending domain
The live toolchain — MCP-driven
Models & tools — Copilot · Gemini · ChatGPT · Claude Code · Cowork Agentic stack — subagents · skills · slash commands · hooks · workflows Atlassian MCP — BRD & FSD authoring · JIRA create / update · Confluence Miro MCP — architecture & flow diagramming Puppeteer MCP — browser automation · QA evidence · UAT walks Microsoft 365 MCP — documents & data Filesystem — repo read / write & navigation Prototyping — pixel-identical, live DOM / CSS extraction HTML docs & decks

One real shipped feature

Data Sense — a worked example, not a slogan.

The rest of this page demonstrates a delivery method. This card grounds the method in a concrete feature: Data Sense, the rule engine behind the Document Intelligence module — designed during my Senior BA tenure and driven to production through my Business Solutions Manager role. First shipped to production in November 2025 on the prior LOS, then migrated and enhanced into Quantum LOS at the February 2026 launch.

Designed
Four-mode automation matrix · Rule taxonomy (Text / Number / Date / Calculate-Date Match) · Action set (Update / Acknowledge / Reject) · Levenshtein text-similarity scoring
Shipped against
Four launch document types — Drivers Licence, Passport, Medicare, Payslip — covering identification + income evidence verification.
Built into
Quantum LOS — the supporting-documents drawer, the rules-administration UI, and the per-rule auto-action gating that controls fully-unattended verification.
Status
●  Live in production since November 2025; migrated into Quantum LOS at the Feb 2026 launch. v2 roadmap (additional doc types + custom rule authoring) is mine to lead as Product Lead.

How Data Sense verifies a document

The rule engine in flight — anchored document to verified source data, attended only when it needs to be.

1
Document anchored
A supporting document is attached against an application requirement (e.g. Payslip → income evidence).
2
Process document
Kicked off automatically by the Processing Policy, or manually by the assessor.
3
Extract & compare API round-trip
Sent out for extraction; fields come back; Data Sense scores extracted values against the captured application data — text / number / date match, Levenshtein similarity.
◇ Per-rule auto-action gate configurable per rule
Auto-pass
✓
Verified, unattended
Within tolerance and auto-action enabled → verified with no human touch.
Manual review · exception
!
Side-by-side comparison
Document renders beside the Data Sense values; mismatches highlighted.
↳
Resolve
Action set — Update · Acknowledge · Reject.
✓
Source data updated · document verified
Outcome written back to the application; verification status and audit trail updated.

The Agentic SDLC

Eleven agents. One auditable pipeline — most runnable standalone.

Stage 0 (a roadmap idea) through Stage 10 (post-deploy). The spine is the canonical chain, enforced by approval gates. Agents marked 🔓 also run standalone — invoked on their own at any time. bug-reviewer and ui-baseline-auditor are conditional entries that join the chain at Build. Click any agent for its brief.

🔓 standalone — own command, run any time 🔗 chain-only ⟂ conditional entry
Build — joined by conditional / standalone entries
🚀 merge + deploy to UAT
🚀 deploy to production
✓ GREEN — terminal
⟳ RED — loops back to bug-reviewer

The longer game — gates that earn their own removal

Every agent ends each run with a 💬 structured feedback prompt and memory candidates. Approved learnings persist to a shared memory the whole team reads, so the agents get measurably better run-over-run. That loop is the engine behind the real roadmap — moving the human approval gates from mandatory to earned:

1
Where it is today
Supervised
A human approval gate sits at every stage. Nothing hands off until a person signs the upstream artefact — the safety net while pass rates mature.
2
The mechanism
Measured
The feedback + memory loop accrues per-stage evidence — refinement acceptance, review-finding validity, QA accuracy, false-positive / false-negative trends — stage by stage.
3
The play
Autonomous hand-offs
As a stage clears a satisfactory pass-rate bar, its human gate lifts — agent hands straight to agent. Done per-gate, evidence-led, and reversible if quality regresses.

Critically: gates lift first where an error is cheap and caught downstream (refinement, test design) and stay longest where it's costly or irreversible (the production deploy). post-deploy-regression-checker remains the always-on backstop regardless. The endgame isn't "no humans" — it's a largely autonomous pipeline with people supervising by exception, spending their judgement where it actually moves the needle.

Anatomy of a ticket

One ticket. The seven core artefacts.

A representative origination ticket walked through the pipeline's canonical path. Each tab is the artefact the agent produces at that stage.

Anatomy of an agent

An agent is a job description with hands.

Each agent is a Markdown file. Frontmatter declares the contract; the body teaches it the job. Below is the production definition for the refinement agent — slightly trimmed.

Each agent targets the model that fits its stage — deep-reasoning models for refinement, architecture and review; faster models for the mechanical stages.
Claude Code is the harness; the model is chosen per agent.