Turning complex business rules into reliable AI systems.
An enterprise healthcare platform that turns spreadsheets, contract documents and existing schedule records into validated, reviewable and traceable changes for downstream reimbursement systems. I architected it and lead its AI and backend engineering.
Architected
The whole platform: six services, AI orchestration, data backbone and keyless CI/CD
Built
Three of the six FastAPI services myself; the team built the other three
With others
The React interface, built with frontend collaborators and AI-assisted development
A healthcare fee schedule describes how covered services are reimbursed: procedures and modifiers, markets, effective and cancellation dates, monetary amounts and the rules that interpret them. Operations teams turn approved business changes into data that downstream reimbursement systems use — so an incorrect or uncertain change affects what providers are paid.
One synthetic schedule, FS-A……holds two kinds of data
Monetary rate record · row-oriented
Explicit amounts
Procedure-level allowed amounts with their modifiers, markets and effective periods. Usually supplied as spreadsheets.
Schedule
FS-A
Market
North
Procedure
PRC-1042
Modifier
26
Setting
Non-facility
Allowed amount
$84.20
Effective
1 Jan 2027
Cancellation
—
Reimbursement configuration · document-derived
Rules that interpret amounts
Percentages, templates, anesthesia factors, lesser-of indicators and professional/technical treatment. Often written in contracts.
Schedule
FS-A
Basis
110% of reference
Lesser-of
Billed charges
Anesthesia factor
52.00
Professional / technical
Split applies
Source
Contract §4.2
Input → requirement → engineering responsibility · hover or tap a row
Spreadsheets
Arrive in different structures, with codes, modifiers and dates to check
RequirementConsistent validation before anything reaches downstream systems
Deterministic intake and validation with row-level issues the user can fix
Contract documents
Rules are written in prose and tables
RequirementReliable structured fields, with uncertainty made visible
Repeatable extraction, AI cross-validation and human review
Existing schedule records
A schedule can have several effective periods
RequirementChange exactly the right period and the right rows
Timeline selection and complete business-row identity
Downstream load
Long-running, and can partially succeed or time out
RequirementKnow what succeeded, what failed and what is uncertain
Row-level correlation, durable status and selective recovery
The aim is not to remove judgment but to place it deliberately: people interpret intent, review what was extracted, approve changes and decide when recovery is safe.
02Requirements → controls
Every control exists for a reason.
Each business requirement became an explicit engineering control. Hover over or tap a requirement to see the control it activates and why it is there.
01
Requirement
Ambiguous intent
Control
Clarification and scenario selection
Requests arrive in plain language and are often incomplete. The intent layer asks for what is missing and resolves a registered scenario — a new schedule, a change to an existing one, or an investigation — before any work begins.
02
Requirement
Invalid business relationships
Control
Deterministic reference checks
Active procedure codes, modifier validity, duplicate code-and-modifier pairs, global versus professional/technical relationships and fee precision are checked by code against reference data. AI may guide the user; it never decides whether a row is valid.
03
Requirement
Precise edits
Control
Complete row identity and constrained fields
An edit is matched on the full business identity of a row, so a neighbour with similar attributes is never changed by mistake. Only approved fields can be filtered, sorted or edited.
FS-A · North · PRC-1042 ·26edited
FS-A · North · PRC-1042 ·TCuntouched
04
Requirement
Business-critical changes
Control
Explicit approval gates
Working data is reviewable and reversible. Nothing reaches the system of record until a person approves it, and a downstream action cannot start before its prerequisites are complete.
05
Requirement
Long-running operations
Control
Persistent workflow state
A shared job context and stage status survive restarts and replica changes, so a long job resumes from what is known instead of starting over.
06
Requirement
Partial delivery
Control
Row-level outcomes and selective recovery
Downstream responses are correlated back to their source rows. Successful rows are preserved; only rows that remain eligible are retried.
07
Requirement
Lost response
Control
Reconcile before retry
A timeout after a state-changing call is an unknown outcome, not a failure. Durable records are checked first, because resending immediately could duplicate work the downstream system already did.
03AI, rules and people
AI handles ambiguity. Code handles consequences.
The architecture deliberately avoids asking a model to perform critical database updates or downstream loads. AI interprets language and documents; deterministic code validates and executes; people correct and approve.
AIinterprets
Intent interpretation
HowGoogle ADK runtime with Vertex AI models
WhyPlain-language requests are ambiguous; the intent layer classifies the scenario and requested action.
Contextual cross-validation
HowModel reasoning over the extracted fields
WhyChecks extracted values against their context and flags low-confidence results for review.
Codevalidates and executes
Document extraction
HowText, table and coordinate reading
WhyLocating expected values must be repeatable for the same document.
Business validation
HowReference and relationship checks
WhyRules must be repeatable and testable — they are never delegated to a model.
Workflow sequencing
HowRegistered playbooks in a deterministic engine
WhyBusiness-critical steps run in a fixed, auditable order and pause at decision points.
Personcorrects and approves
Correction and approval
HowReview workspace
WhyConsequential changes need a person’s decision, and approval is separate from editing.
Phase 1 · agentic edge
Intent and context
An ADK-based intent layer interprets the request, keeps conversation context and resolves the target scenario. Sessions are persisted, so a multi-turn conversation survives a restart or a move between replicas.
Phase 2 · deterministic core
Playbook execution
A deterministic engine advances step by step, calls registered capabilities and waits at explicit human decision points. MCP is used as a connector boundary to downstream tools — not as the business workflow.
Simulated reviewSynthetic contract
One extracted field, end to end
Source context · §4.2“Reimbursement for covered services shall be the lesser of billed charges or 110% of the reference schedule, except where Exhibit B applies.”
CodeExtraction
AICross-check
CodeValidation
PersonReview decision
✓Approved
Basis 110% of reference · lesser-of billed charges
Lesser-of condition depends on Exhibit B · low confidence
Percentage within policy range ✓ · lesser-of ⚑ needs review
Waiting for a reviewer. Approval stays closed.
Reviewer corrected it: the Exhibit B exception applies. Revalidating…
Corrected by reviewer · revalidated ✓ · ready for approval.
Approved by the reviewer. Only now can the configuration move toward delivery.
A decision is revalidated before approval becomes available — editing never authorises a change on its own.
04Architecture
Explore the system, one responsibility at a time.
Hover over or tap any part of the platform to see its responsibility, the requirement it satisfies, its technology and its safeguards. The connected path lights up while the whole system stays in view.
Experience
Orchestration
Domain capabilities
Data, state and the system of record
Cross-cutting
Guided path Focused direct action
05Data and technology
Files, analytics and operational state are not interchangeable.
Each store has one job. A shared job context travels through all of them, so a support engineer can connect a UI action to a file, a database state, a service log and a downstream result.
Files
Cloud Storage
Source files, corrected outputs, exports and downloadable evidence.
Structured and analytical
BigQuery
Reference data, job staging, analytical views and workflow records.
Operational state
Cloud SQL (MySQL)
Persistent agent sessions and operational tables for explicitly enabled paths.
Persisted session · Cloud SQLJob context and conversation intact
Data lifecycle
source
normalised working data
approved records
downstream delivery
reconciliation evidence
A controlled database boundary
Operational persistence was introduced behind compatibility boundaries rather than as a wholesale replacement. Selected preparation and delivery paths can use Cloud SQL through a strict configuration gate, while BigQuery keeps reference data, analytics, broad staging and the flows that have not moved.
A request never mixes BigQuery and MySQL operations mid-flow.
Database and table identifiers are validated; business values are bound as parameters.
Connection settings and credentials come from the environment and secrets, never source code.
Business contracts are preserved, so the UI and orchestrator need no per-backend workflows.
Technology, by responsibility · hover or tap a technology
Interface
Services
AI
Integration
Data
Platform
Delivery
Operations
Every technology has one job in the platform.
React · TypeScriptInterface · Operator experience
One journey that renders chat, forms, timelines, grids and results from a shared contract
Python · FastAPIServices · Workflow and domain APIs
Typed, testable services with clear business-named capabilities
DockerServices · Packaging
One versioned image per service
Vertex AIAI · Model access
Managed enterprise model execution for intent understanding and document cross-validation
Google ADKAI · Intent-agent runtime
Standard agent lifecycle, model callbacks and persistent sessions
MCP · RESTIntegration · Downstream connectors
A structured tool boundary without embedding every transport detail in the orchestrator
BigQueryData · Reference, staging, analytics, workflow records
Large structured datasets and SQL-based comparison
Cloud SQL (MySQL)Data · Sessions and enabled operational paths
Transactional access and multi-turn state
Cloud StorageData · Files and evidence
Durable objects with controlled access
GKEPlatform · Service runtime
Health checks, rolling updates, autoscaling and workload isolation
Artifact RegistryPlatform · Container images
Versioned delivery to each environment
GitHub ActionsDelivery · Pipelines
Independent build, test, scan and deploy per service
Keyless, scoped authentication — no stored CI keys
Managed Prometheus · monitoringOperations · Service and queue metrics
Autoscaling signals and runtime visibility
Centralised log searchOperations · Investigation
Correlated events across every service
06Interface contract
One conversation, several workspaces.
The interface sends one consistent interaction envelope — the session, the user action and the available context. The orchestrator replies with a message plus structured interaction data that tells the interface what to render next.
Conversation
Upload
Dynamic form
Timeline selection
Review grid
Approval
Results and recovery
Load the new North market rates for schedule FS-A.
Is this a new schedule, or a change to an existing one?
A change — effective next January.
rates_north_2027.xlsxUploaded · 2 sheets · bound to job J-0001
1 Jan 2025 – 31 Dec 20251 Jan 2026 – 31 Dec 20261 Jan 2027 – open
PRC-104226$84.20✓
PRC-1188QX$41.00⚑ invalid modifier
PRC-2001—$12.75✓
3 rows ready · 0 open issues
Approve and deliver
2 delivered1 needs correctionRetry eligible rows
Synthetic interface mock-up
01 / 07
Conversation
Intent discovery, clarification and status questions
Behind the screenIntent classified as a change to an existing schedule · missing context requested · playbook selected
02 / 07
Upload
The selected playbook requires a spreadsheet or document
Behind the screenSource file stored and bound to the job context · intake stage started
03 / 07
Dynamic form
Structured fields cannot be inferred safely
Behind the screenWorkflow paused at a required-input step · fields defined by the playbook
04 / 07
Timeline selection
An existing schedule has several effective periods
Behind the screenMatching periods retrieved · user’s selection fixes the target period for every edit
05 / 07
Review grid
Validation returns row-level issues the user can correct
Behind the screenValidation results returned · corrections saved and the same checks rerun
06 / 07
Approval
Business-critical changes need an explicit decision
Behind the screenApproval recorded · stage gate opens for downstream delivery
07 / 07
Results and recovery
A delivery, retrigger or investigation produces an outcome
Behind the screenRow-level outcomes persisted · only eligible rows offered for retrigger
Guided orchestration
Most guided interactions return through the orchestrator, so the conversation, the workflow step and the interface state stay synchronised.
Focused direct actions
Some focused panels call a service directly when the action is already unambiguous — such as paging or editing a review grid. Not every click passes through the orchestrator.
I defined the interaction contract and architected the backend behind it. The React interface was built with frontend collaborators and AI-assisted development.
07Delivery and recovery
The last mile is a correlated batch, not a blind write.
Approved records are grouped by business context, sent in bounded batches and correlated back to their source rows. Keep scrolling to see what is known, what is uncertain and what is safe to do next.
Selecteligible or failed rows
Groupschedule and market context
Batchbounded request envelopes
DeliverREST or MCP transport
Correlaterow-level responses
Persiststatus, audit and summary
SyntheticEight rows in two batches
Explicit rejection
Partial completion
Timeout or lost response
Completed with evidence
Batch 1
Row 1PRC-1042 · 26✓ Success
Row 2PRC-1042 · TC✓ Success
Row 3PRC-1188 · —✓ Success
Row 4PRC-2001 · —✓ Success
Batch 2
Row 5PRC-2310 · 26✓ Success
Row 6PRC-2310 · TC✕ Failed
Row 7PRC-3105 · —✓ Success
Row 8PRC-4020 · —✓ Success
Explicit rejection
Request rejected with a recorded error. Row 6 rejected: modifier not valid for this procedure. Seven rows keep their durable success.
Evidence established
✓Transport — a response was received
✓Service — the service finished its action
✓Persistence — durable rows were written
·Business — the UI and downstream system show the intended outcome
Safe next action
Fix the rejected row and retrigger only that row.
Explicit rejection
Safe action taken
Fix the rejected row and retrigger only that row.
Row 6 corrected and retriggered alone → ✓. Nothing else was resent.
Partial completion
Some rows have durable success and others do not. Batch 2 stopped after its first row. Rows 1–5 are complete and are not resent.
Evidence established
✓Transport — a response was received
✓Service — the service finished its action
✓Persistence — durable rows were written
·Business — the UI and downstream system show the intended outcome
Safe next action
Retry only the eligible non-success rows.
Partial completion
Safe action taken
Retry only the eligible non-success rows.
Rows 6–8 retried → ✓. Rows 1–5 were preserved, not repeated.
Timeout or lost response
The outcome is not proven either way. Batch 2 was sent but no response arrived. Rows 5–8 are uncertain — not failed.
Evidence established
✓Transport — a response was received
·Service — the service finished its action
·Persistence — durable rows were written
·Business — the UI and downstream system show the intended outcome
Safe next action
Reconcile against durable records before any retry.
Timeout or lost response
Safe action taken
Reconcile against durable records before any retry.
Reconciliation found rows 5–7 already applied → marked ✓ without resending. Row 8 was not applied → eligible for a retry.
Completed with evidence
Source, audit and workflow state agree. Every row has a correlated success, durable status and a matching downstream outcome.
Evidence established
✓Transport — a response was received
✓Service — the service finished its action
✓Persistence — durable rows were written
✓Business — the UI and downstream system show the intended outcome
Safe next action
Present the final result to the user.
Four kinds of evidence
A successful HTTP response proves transport. A healthy service proves it completed its own action. A durable write proves persistence. Only the interface and the downstream system agreeing proves the business outcome — final status comes from durable workflow evidence, never from a status code.
Recovery reuses work
Document interpretation persists its stage output, so a resumed job reuses it instead of repeating expensive extraction. Termination-before-add flows resume from durable stage state.
08Production engineering
From reviewed change to verified service.
Six independently deployable services share one delivery pattern across four environments, with the same controls everywhere.
01Reviewed changetargets the right delivery branch
02Tests and quality checksautomated tests and code-quality scanning
03Versioned containerone image per service
04Artifact publicationapproved images stored in registries
06Operational verificationhealth, rollout, metrics and business evidence
01
Independent service delivery
Separate pipelines reduce coupling; common controls keep consistency.
02
Environment-specific configuration
Development, test, stage and production use separate namespaces and settings.
03
Keyless access
Pipelines authenticate through Workload Identity Federation — no stored CI keys.
04
Runtime secrets
Injected at runtime, never kept in source or deployment text.
05
Least privilege
Environment-specific service accounts with scoped cloud permissions.
06
Health probes
Readiness and liveness checks protect traffic and enable self-healing.
07
Scaling signals
Request and queue metrics drive replica changes from observed demand.
08
Correlated, bounded logs
Job, session, service and correlation context; allow-listed identifiers only — no request bodies or credentials.
09
Business completion evidence
A healthy pod confirms delivery readiness; it does not prove the rate change happened.
Quality gate as stated in the pipeline: at least 90% test coverage, with static analysis, before a service can be packaged. This is a stated gate, not a measured business result.
09My contribution
What I did, and what the system does.
Personal contribution and documented platform capability are kept separate. Contribution statements come from my own record of the role; capabilities come from the platform’s architecture documentation.
My contribution
Architected the end-to-end platform and led its AI and backend engineering.
Designed all six independent FastAPI services on GKE with horizontal pod autoscaling; built three of them myself, and the team built the other three.
Designed the use of Vertex AI for PDF extraction, confidence signals and natural-language intent routing through a conversational orchestrator.
Designed the BigQuery staging and audit layer and Cloud Storage file handling, with a human sign-off gate before any authoritative write.
Designed keyless CI/CD with GitHub Actions and Workload Identity Federation, static analysis and coverage gates across four GKE environments.
Defined the interaction contract the React interface renders from; the interface itself was built with collaborators and AI-assisted development.
Documented platform capabilities
Deterministic intake validation with row-level issues and revalidation.
Separate handling of monetary records and document-derived configuration.
Conversational intent handling around deterministic playbooks.
Bounded, correlated downstream delivery with selective recovery.
Investigation views with workflow status, row evidence and stable error categories.
Independent services with shared identifiers across one business journey.
No business metrics are claimed: the platform’s value is described qualitatively — more consistent rate quality, fewer manual handoffs, faster issue resolution and safer recovery.