Azure Integration & AI Platform Architect HI, I'M AKSHAY Kakoriya
I design API-first, event-driven integration on Azure — and lately, the AI systems that sit on top of it.
Gwalior, India · UTC+5:30 · Remote
- AZ-305
- AZ-204
- CLAUDE ARCHITECT
- 7 YRS
- AZURE
- interfaces built two estates, 2022 — 25
- 60+
- integrations designed one estate of 200+, 2025 — 26
- 50+
- fewer recurring incidents one estate, 2022 — 24
- ~30%
- lower Azure spend one estate, 2022 — 24
- ~15%
Introduction
About.
I design enterprise integration on Azure — API-first, event-driven systems that connect internal platforms to the outside world without breaking under load. Lately I've been doing the same for AI systems, and most of the hard parts turn out to be the same ones.
Seven years across Accenture, Cognizant and Capgemini took me from building interfaces to owning the architecture behind them: requirement analysis, design authority, solution design, and the production behaviour that decides whether a design was actually any good. Most recently that meant an ERP integration layer delivered against a fixed date, and the target architecture for moving a two-hundred-integration estate off its data centre.
The work I'm proudest of is rarely on the happy path. It's the retry policy, the dead-letter route and the idempotent handler that decide whether an integration survives a dependency failing — and the instrumentation that tells you which one fired. Most recently that same thinking carried into agentic AI on Azure AI Foundry, where what mattered was boundaries and evaluation, not model choice.
Integration Architecture
API contracts written and versioned before the implementation, policy-enforced at the boundary.
Case studyEvent-Driven Systems
Asynchronous by default, synchronous by exception — and what that costs when it is wrong.
The decisionsAgentic AI on Azure
Agent boundaries, retrieval isolation and evaluation on AI Foundry — a build of my own.
The projectReliability Engineering
Retry, dead-lettering and idempotency, written down as patterns rather than rediscovered.
The patternsWhat I've built
Selected work.
Four engagements, newest first, then a project of my own and one practice page — each written around the decisions rather than the outcome: what was considered, what was chosen, and what that choice cost.
Two hundred integrations, counted before designed
A legacy integration estate with a data-centre exit date and no inventory — the count that had to come first, the four patterns it produced, and the design that was approved as the baseline for the build.
- integrations inventoried
- 200+
- assigned a target pattern
- 50+
- low-level designs written
- 20+
Thirty interfaces, one go-live date
Bespoke on-premises applications wired into a new ERP against a fixed date, with requirements still moving — the contract-first discipline that absorbed the churn, and what a batch of messages is worth.
- Service Bus messages, peak day
- 500K+
- interfaces live at go-live
- 30+
- new-interface delivery
- 3 wks → 1 wk
An estate nobody had documented
An integration estate taken over undocumented, deployed by hand and failing in ways nobody could explain — the order the team took it in, and the five decisions inside it that were mine.
- fewer incidents, estate-wide
- ~30%
- lower Azure spend, estate-wide
- ~15%
- interfaces built or changed
- 30+
The day the platform was built for
A utility's public customer platform meeting its first real storm — what gave way, what the team changed afterwards, and which of those changes were mine.
- services and jobs shipped
- 20+
- API calls a day, normal load
- 100K+
- environments, released per service
- 3+
Retrieval that cannot leak, and an upgrade refused
My own build: a retrieval system over 1.5 million chunks where source isolation is a property of the topology rather than a prompt instruction, and where an 87-case evaluation set talked me out of a model upgrade.
- chunks indexed
- 1.5M
- groundedness, 87 cases
- 4.32/5
- the upgrade refused
- 19.7% vs 16%
Reliability patterns for enterprise integration
The four patterns that decide whether an integration survives production — what each one costs, and when not to reach for it.
- Retry with backoff
- Dead-lettering
- Circuit breaking
- Idempotent processing
Capabilities
What I work with.
Azure Integration
- Logic Apps
- API Management
- Azure Functions
- App Service
- Service Bus
- Event Grid
- Integration Account
- Key Vault
- Azure Cache for Redis
- Storage Accounts
- Azure SQL
- Front Door
- WAF
AI on Azure
- Azure AI Foundry
- Agent Framework
- Multi-agent orchestration
- Tool calling
- Azure OpenAI
- Azure AI Search
- Embeddings
- LLM evaluation
- Prompt-injection guardrails
Architecture & Delivery
- Technical Design Architecture
- High availability
- Fault tolerance
- Solution estimation
- Cost optimisation
- Agile Scrum
Security & Observability
- Managed Identity
- Microsoft Entra ID / MSAL
- OAuth 2.0
- RBAC
- Application Insights
- Log Analytics
- KQL
- Azure Monitor alerting
DevOps & IaC
- Azure DevOps YAML
- Bicep
- ARM templates
- Terraform
- CI/CD automation
- Multi-environment promotion
- Git branching strategy
- pytest
Integration & B2B
- EDI / EDIFACT
- AS2
- Liquid
- REST / JSON
- XML
- OData (Dynamics 365 F&O, SAP S/4HANA)
- Salesforce REST API
- OpenAPI
- Event-driven patterns
- API-first design
Languages
- C# / .NET
- Python
- SQL
What I've done so far
Experience.
-
Nov 2024 — Oct 2026
Architecture
Senior Consultant
CapgeminiClients: an Australian construction, property and finance group; a European ferry and cruise operator.
- Technical lead on the integration layer wiring bespoke on-premises applications into Dynamics 365 Finance and Operations: designed the chain every interface is built on, authored most of the design documents, and led and reviewed two to three developers.
- Built and took 30+ interfaces live on a single ERP go-live date — API Management to Service Bus, then Logic Apps Standard and Functions mapping in C# and Liquid into data management packages or OData — with zero defects raised in user acceptance testing.
- Cut import runs by batching Service Bus messages into one data management package per window, and kept a slow or unavailable ERP off the user by acknowledging with a correlation ID and delivering behind the queue. Peak day: 500K+ Service Bus messages.
- Designed for partial failure: bounded retry with backoff and timeouts, a dead-letter route per interface alerting a named owner, business-key upserts and Service Bus duplicate detection, and a feedback loop that surfaced failed package rows individually.
- Built the delivery machinery from day one — Azure DevOps YAML with Bicep across Dev, Test and Prod, Application Insights alerting — and cut new-interface delivery from ~3 weeks to ~1 week with reusable workflow, policy and pipeline templates.
- Led the integration architecture for a ferry operator leaving its data centre: inventoried and classified its 200+ integration estate of Java, Camel, MQ and file transfers, set four target patterns on Azure Integration Services, and wrote 20+ approved designs.
- Architected and built a multi-agent emergency-response solution on Azure AI Foundry for the 2026 hackathon, in a team of three. Mentored four developers across the two engagements.
- API Management
- Service Bus
- Logic Apps Standard
- Azure Functions
- Integration Account
- Dynamics 365 F&O
- Bicep
- Azure DevOps
- Azure AI Foundry
-
Aug 2022 — Nov 2024
Ownership
Associate — Projects
CognizantClient: a global 4x4 vehicle manufacturer (UK) — the integration estate behind its public customer website, taken over from a previous vendor by the team I joined.
- Built or materially changed 30+ interfaces behind the website's configure, reserve and order journeys, from requirement analysis and design review to go-live and L3 support: APIM and .NET Functions to SAP S/4HANA over OData; Service Bus to Salesforce REST.
- Redesigned the error handling: bounded retry, backoff and timeouts on every downstream call, a dead-letter route per interface with a named owner and runbook, correlation IDs end to end, idempotent writes — my part of a team-wide ~30% fall in recurring incidents.
- Took SAP and Salesforce off the website's critical path after a slow downstream became a slow website: Azure Cache for Redis on SAP reference data, Salesforce writes made asynchronous over Service Bus, Event Grid blob-event triggers for the file-exchange flows.
- Replaced secrets with identity after an expired key took an interface down: OAuth 2.0 on APIM (validate-jwt against Entra ID; a client-credentials app for the website vendor) and managed identity APIM → Functions and Functions → Key Vault, Service Bus, Storage.
- Moved most of the estate's function apps from in-process .NET 6 to the isolated worker on .NET 8; consolidated 10–20 Consumption Logic Apps onto one Standard plan and stepped premium tiers down — my share of a team-wide ~15% cut in Azure spend.
- Built most of the team's YAML pipelines — Bicep and zip deploy, one artefact promoted through Dev, Test and Prod — replacing manual portal deployments; built the KQL alerts that reach a named owner on failures and dead-letter depth; mentored 3–4 new joiners.
- API Management
- Azure Functions
- Logic Apps
- Service Bus
- Azure Cache for Redis
- Event Grid
- Key Vault
- Managed Identity
- Azure DevOps
- C# / .NET 8
-
Jun 2021 — Aug 2022
Production
Application Development Analyst
AccentureClient: a US energy utility — its public customer platform.
- Owned named backend services end to end — design, build, per-service release and production support on a follow-the-sun rotation — behind API Management for the utility's public customer site and contact-centre tools.
- Built services in the monolith-to-microservices split — 10–20 services, boundaries drawn by backend dependency — and the later trigger-driven services on Azure Functions.
- After a storm-driven surge maxed out App Service and tripped APIM throttling, designed the CPU-based autoscale rules and the gateway response caching for reference reads; recognised with a Sparkling Star Award (2021).
- Wrote the per-service runbooks — deployment and incident response — that the support team adopted.
- App Service
- Azure Functions
- API Management
- Azure SQL
- Application Insights
- Azure DevOps
- C# / .NET
-
Jun 2019 — Jun 2021
Foundations
Application Development Associate
AccentureClient: a US energy utility — its public customer platform.
- Built C#/.NET backend APIs and WebJobs on Azure App Service, exposed through API Management to a CMS frontend and IVR tools — usage, outage and service-request journeys at hundreds of thousands of API calls a day.
- Consumed client-owned SOAP services fronting a legacy mainframe for account verification and usage data, with SQL Server on Azure as the platform's system of record.
- Wrote the WebJobs that purged aged outage and audit data, holding database size, cost and query latency in check as the platform grew.
- Extended the team's Azure DevOps pipelines with automated test stages and per-service pipelines; completed Accenture's Full-Stack Automation Engineer learning path (Feb 2021).
- App Service
- WebJobs
- API Management
- Azure SQL
- SOAP / XML
- Azure DevOps
- C# / .NET
Thinking out loud
Writing.
- 01 3 min read
Designing for the third party having a bad day
The reliability patterns that decide whether an integration survives production
Read - 02 5 min read
Agents that dispatch help
A multi-agent roadside-emergency platform on Azure AI Foundry — and why the agent boundaries sit where they do
Read - 03 3 min read
Making isolation structural, not conversational
Retrieval architecture for a RAG system over regulated material, and why the guardrail lives in the topology
Read - 04 4 min read
The model upgrade I didn't do
Why a newer model made my RAG system worse, and what the evaluation actually told me
Read -
More writing
New writing lives at kakoriyacloud.com
New posts go up there — integration, reliability, and the AI systems now being built on top of them. The essays here stay here.
Visit kakoriyacloud.com
Certifications & education
Credentials.
Microsoft
Anthropic
SCRUMstudy
- SFC
Scrum Fundamentals Certified
May 2021
Education
Bachelor of Engineering, Computer Science
Madhav Institute of Technology & Science (MITS) · Gwalior, Madhya Pradesh
2015 — 2019
Languages
- English
- Full professional (C2)
- Hindi
- Native
Honours
-
Capgemini Hackathon 2026 — with Microsoft
A multi-agent emergency-response solution on Azure AI Foundry, architected and largely built by me in a team of three over about four weeks: three or four agents on GPT-4o taking a chat intake into a structured incident record, triaging severity against a rubric, and dispatching a responder — calling simulated geolocation, vehicle telemetry and nearest-facility services as tools.
-
Sparkling Star Award — Accenture
For the storm-surge response on a utility customer platform — the autoscale rules and the gateway caching that followed. 2021.
Get in touch
Contact.
Open to Azure integration and AI platform architecture work — permanent or contract. Based in Gwalior, India (UTC+5:30), working remotely, as I have with teams in the UK, Australia and the US. If you have an integration estate that needs an owner, or a design you want a second opinion on, email is the fastest way to reach me.