Navalo
Menu

I'm Navalo. I build Workday integrations, and the AI agents that build them.

By day I deliver Workday integrations, as I have for eight years at Alight and Cognizant. After hours I taught myself AI engineering and built a team that never sleeps: WICKY, ten agents that take an integration from the first requirements question to a tested deployment. They show their working, and they never skip the documentation.

See the agentsMy experience
  • 10 agents on one platform, WICKY
  • 130 tools across five working MCP servers
  • 3 agents that also ship as products of their own
  • 1 developer. That's me, in my own time.

What I've built

Everything here is a personal project, designed and built on my own time. Each agent encodes something I learned the hard way in delivery, and each one refuses what it can't verify rather than produce something that only looks finished.

WICKY

Workday engineering platform

Ten agents share one memory and carry a Workday integration change from an incident or request through discovery, design, build, testing and deployment, then learn from the result. It handles new integrations and changes to existing ones, and versions everything forward instead of overwriting.

Every handoff between agents is validated against a shared contract. A governance agent stops personal data before it moves, and the delivery console keeps integrations, tasks and approvals in front, with the agents behind the scenes. The full loop has run end to end on a live system.

Agents that show their working

Every agent is built to prove its work, and a person approves anything that touches a tenant. Each walkthrough plays once when it scrolls into view, and you can click any step to jump to it. The data in the examples is made up; the steps are what the agents actually do.

Thirattu

Requirements discovery

Runs the requirements interview for a new integration. It starts from the business outcome, asks only the questions that matter for that kind of integration, and turns the answers into the documents a build needs.

MCP server with 14 tools, TypeScript

Proven: all 18 of its automated tests pass on the current code (November 2026). They start the real server and play a complete interview through it, from the first question to a finished spec document, XSLT starting point and UAT matrix.

Send benefit enrolments to our dental carrier every week.Classified as benefits, outbound, scheduledFull file.Next question chosen from your answers, not by a model
Open item: carrier file specification not received yet

What does this integration need to do?

Full file every time, or only changes since the last run?

Which dependants should the carrier receive?

Integration spec
XSLT starting point
UAT matrix
Illustrative example

Before

Requirements are gathered across workshops and emails, and the gaps show up halfway through the build.

With Thirattu

Gaps are listed on day one, and the spec, XSLT starting point and test matrix come out of the same interview.

Goddy

Workday Studio builder

Builds Workday Studio integrations from a filled-in intake. It writes the assembly, checks it against Studio's own rules, repairs known defects and reviews the result the way a senior developer would, then packages it for import.

MCP server with 38 tools, Node.js. Runs standalone or inside WICKY.

Proven: all 149 of its automated tests pass on the current code (November 2026). They start the real server, build its reference integrations into packages that come back clean from validation and senior review, and confirm that a build with a planted production credential is refused.

Intake
Direction
Outbound
Source
Custom report
Destination
File to carrier SFTP
Trigger
Weekly schedule
Checks

Assembly matches Studio's schema

Web service request built from a known pattern

Delivery step has no error handler

Error handler added, checks re-run

Senior review: the report runs once per employee inside a loop. Run it once and look each employee up instead.

Assembly
Launch on weekly schedule
Get data from custom report
Transform with XSLT
Deliver file to carrier SFTP
Error handler and notification
Dental carrier outbound packaged for import into Studio, with its route contract
Illustrative example

Before

A Studio build starts from a blank assembly and whatever the developer remembers about what went wrong last time.

With Goddy

It starts from a draft that's already validated, repaired and reviewed against rules learned from real builds.

Kevin

Orchestrate and Extend builder

Plans and builds Workday Orchestrate integrations and Extend apps, and converts older integrations. Every decision it makes comes with a route contract, so a reviewer can see what it chose, what policy allowed and why.

MCP server with 25 tools and more than 1,500 automated tests, Node.js

Proven: on the current code, 1,570 automated tests pass and none fail (November 2026). They start the real server and call its tools, and check that every built-in template builds into a valid orchestration file. Another 20 are skipped because they need data that isn't kept in the repository.

"When someone is hired, create their badge in the building access system."

Orchestration
Hire business process completes
Get worker details from Workday
Map name, ID and work location
Create badge in access system
Record the result
Checks

Valid orchestration format

Built from a verified example

Error path and retries present

Route contract
CONSTRAIN

Approval required: it sends personal data to an outside system.

Delivery stays disabled until a person approves.

Badge on hire packaged, with an evidence pack and the route contract
Illustrative example

Before

When AI generates an integration, the reason it chose that path is buried in a chat log, if it's anywhere.

With Kevin

Every decision carries a contract a reviewer can read, and anything that writes personal data waits for a person.

RANJ

Release regression agent

Tests integrations before and after a Workday release. It compares the new output with a known-good baseline field by field and reports which changes are real. It can also check a single file against its spec before handoff.

MCP server with 38 tools, Python. 19 file checks for single-file validation. Refuses to run against production.

Proven: it has run its full cycle on a live system, not only against mocks. It relaunched an integration with the baseline run's settings, waited for it to finish, fetched the new file and compared the two. That run also exposed two defects the mocked tests had missed, and both are fixed.

FieldBaseline, before releaseAfter release
Employee ID1048210482
PlanDental PPODental PPO
Coverage start01/01/202601/01/2026
Coverage end12/31/2026(blank)
File created03/02 09:0003/09 09:00 expected
1 change to review1 expected difference set aside3 fields unchanged

Vendor delivery restored

Illustrative example

Before

Release testing means opening two files side by side and checking them by eye, twice a year, for every integration.

With RANJ

Every field is compared, and the report says exactly which changes need a person's attention.

Hattori

Supervised tenant configuration

A browser agent that configures a Workday implementation tenant: calculated fields, custom reports, EIBs, Core Connectors and business process definitions. Each change follows a typed recipe, waits for a named approval and saves evidence. It never launches or activates an integration.

Chrome extension, local bridge and MCP server with 15 tools

Proven: in one approved run, it built a calculated field, a report that uses the field and an outbound EIB that uses the report. Two preview rows were checked against the source records and both matched. The EIB's event history shows zero runs, because nothing was launched.

Recipe: calculated field

Field name given

Business object given

Function and inputs given

Approval

Create 1 calculated field in the implementation tenant

Approved by Navalo
Implementation tenant
Create Calculated Field
Years of Service
Worker
Date Difference
Hire Date to today
Evidence saved screenshot, field values and timestamp. Nothing launched or activated.
Illustrative example

Before

Tenant configuration is clicked through by hand and documented afterwards, if it's documented at all.

With Hattori

Each change is planned, approved by name, carried out and left with its own evidence.

How I work

Integrations move pay, benefits and personal data. These are the rules every tool I build follows.

Finished is not succeeded

A run that completes can still send the wrong file. Every test compares against a known-good baseline, and choosing that baseline matters more than the run itself.

The model proposes, code decides

A language model can reason about what an integration needs, but the build, validation and packaging steps are deterministic and can be replayed. The model never grades its own work.

A person approves every tenant write

Reading is automatic. Anything that changes a tenant waits for a named approval and leaves evidence behind.

Say what's proven, and how

Every claim is either measured, sourced or labelled as inference. When the tools can't ground an answer, they say so instead of filling the gap.

{
  "routing_lifecycle": {
    "intent": "convert_integration",
    "risk_score": 0.62,
    "validation_plan": "sandbox-test"
  },
  "policy_governance": {
    "policy_verdict": "CONSTRAIN",
    "approval_required": true,
    "constraints_applied": ["delivery-disabled"],
    "geographic_boundary": "data-residency-review"
  },
  "telemetry": {
    "audit_trace_id": "…",
    "confidence": 0.8
  }
}
ALLOWCONSTRAINDENY

A route contract, from my Route Engineering paper: one portable record of what an AI system decided, what policy allowed and how to audit it. Kevin and Goddy attach one to every decision.

Experience

Eight years delivering Workday integrations for clients across APAC, EMEA and North America, in implementation projects and in production support.

Sep 2018 to Oct 2023

Alight Solutions

Workday Analyst and Integration Consultant, Chennai

5years
2releases a year
0client escalations
  • Delivered Workday integrations for global clients with Studio, EIB, Core Connectors, PECI, PICOF, benefit templates, reports and calculated fields.
  • Built and supported benefit carrier integrations for major US vision, dental, life and health insurers, including eligibility rules and carrier-specific file layouts.
  • Supported Workday's twice-yearly releases with regression packs and test evidence, with no client escalations.
  • Received the Annual Platinum Team Award for supporting multiple AMS clients.
Nov 2023 to now

Cognizant Technology Solutions

GDC Analyst Consultant, Workday Integrations, Chennai

2implementations at once
9person AMS team
24/7support cycle
  • Sole integration developer on two Workday implementations running at the same time, for a chemicals manufacturer and a city government, owning each integration from requirements to deployment.
  • Led a nine-person AMS team on round-the-clock support for a global pharmaceutical company new to Workday, taking it from hypercare to stable operations across payroll, benefits and HCM.
  • Builds with Studio, EIB, Core Connectors, document transformation and Orchestrate, using XSLT and SOAP and REST web services.
  • Presented AI and automation prototypes to leadership, including a working demo of RANJ for integration regression testing.
Credits

Certifications

  • Workday Pro Integrations Core
  • Workday Studio for Integration Consultants
  • Workday Orchestration
  • Workday HCM Essential Concepts
  • Anthropic: Claude Code in Action, Claude 101

Industries

  • Pharmaceuticals and life sciences
  • Chemicals and manufacturing
  • Public sector and local government
  • Healthcare, insurance and benefits
  • Cybersecurity and IT

Education

  • Bachelor of Computer Applications
  • Guru Nanak College of Arts and Science, Chennai, 2018

How I taught myself

Nobody assigned this work. I learned AI engineering by building, one problem from my own delivery work at a time.

  1. 2025

    First prototypes

    An assistant that explains integration errors in plain language, a change-request assistant, and n8n workflows with GPT and Gemini for support tickets.

  2. June 2026

    Agents as tools

    Goddy, Kevin and RANJ became MCP servers that AI coding tools can call, with their knowledge served from the server itself.

  3. July 2026

    One platform

    WICKY brought the agents together on shared contracts, with a governance agent, a delivery console and a closed loop from incident to deployment. Thirattu was rebuilt around business questions.

  4. August 2026

    Into the browser

    Hattori started doing configuration work through a supervised browser, with approval and evidence for every step.

  5. September 2026 on

    Standalone products

    Goddy, Kevin and RANJ each ship as their own product, built inside WICKY and published only after their tests pass on a fresh clone.

What I know first-hand

The agents encode what I've learned delivering Workday integrations.

Integration builds

  • Workday Studio
  • Workday Orchestrate
  • Workday Extend
  • EIB and document transformation
  • Core Connectors and Cloud Connect for Benefits

Data and interfaces

  • SOAP and REST web services, JSON
  • Reports as a service
  • PECI, PICOF and WECI
  • Benefit carrier feeds
  • XML, XSLT 2.0 and 3.0, XPath

Delivery

  • Requirements and integration design
  • Calculated fields and custom reports
  • Tenant migration and deployment
  • UAT and release regression testing
  • AMS leadership with ServiceNow and Jira

Agent engineering

  • MCP servers in Node, TypeScript and Python
  • Multi-agent systems on shared contracts
  • Deterministic validation and self-repair
  • Supervised browser agents
  • Claude, GPT and Gemini, n8n

Also

Route Engineering

Whitepaper, v0.2.0, public domain

Every time an AI system picks a path, it makes a routing decision whose reasons usually live in prose. The paper defines one portable contract for the decision, the policy verdict and the audit trail.

CORTEX

Agent development framework

One set of agent specs that compiles to AGENTS.md, Claude Code or a plain system prompt, so the same rules work with any model.

Ningle

Phone-first social app

A camera and chat app where every friendship shares a virtual pet that grows with your streaks. Built with React 19, TypeScript, Tailwind and Supabase.