Vetri Krishnaraj

Principal Software Architect · Scottsdale, Arizona · Remote

I build production AI inside systems that were never designed for it.

Twenty years of enterprise architecture, the last two spent on applied AI — agentic systems, LLM pipelines, and computer vision — in an environment with real compliance boundaries, real legacy debt, and real consequences for being wrong.

See the work LinkedIn Email

Open to AI architecture and AI engineering leadership roles.

01

Selected work

Agentic systems Prototype

A governed agentic engine for legacy modernization

An orchestrated multi-agent system that rebuilds features of a twenty-year-old enterprise platform from behavior rather than from source code.

How it works Close
  • Clean-room approach: the input is a behavior specification extracted from the running system, not the legacy source. Old code resolves ambiguity; it is never the thing being transformed.
  • Orchestrator/worker model tiering on Amazon Bedrock, running inside the VPC with no egress path for source or data.
  • Automated evaluation harness using the legacy system as a golden master, so the loop iterates against its own failures rather than routing every correction through a human.
  • Governance as architecture: full audit logging, configurable human approval gates, and a per-run token budget with a hard kill-switch.
  • Designed so the engine — not any single rebuilt feature — is the deliverable. Success is measured by how much of the first slice generalizes to the second.
  • Claude
  • Amazon Bedrock
  • AWS VPC
  • Python
  • Java
Computer vision Delivered

Visual identification for unclaimed baggage

A multi-stage vision pipeline that classifies the contents of unclaimed bags to support matching and resale.

How it works Close
  • Dataset curation, labeling, and model training and versioning on Amazon Rekognition Custom Labels.
  • Per-category confidence-threshold calibration across more than forty categories — the difference between a model that scores well and one that behaves sensibly in production.
  • Multi-stage fallback cascade: custom model, then general label detection, then OCR-based text extraction, with domain logic resolving the categories none of them get right alone.
  • Evaluated on F1, precision, and recall per label rather than on aggregate accuracy.
  • Amazon Rekognition
  • AWS Lambda
  • S3
  • Java
LLM systems Delivered

Call transcript analysis at scale

LLM-based analysis of more than 1,200 customer service call transcripts, turned into findings a business leader could act on.

How it works Close
  • Automated classification of calls by type and business unit from transcript content alone.
  • Structured extraction of agent-performance and process-failure patterns across the corpus.
  • Delivered as an executive report to a business unit general manager — the output was a decision, not a dashboard.
  • LLM pipelines
  • Python
  • AWS
Developer platform Ongoing

An internal AI engineering platform

The tooling layer that made everything above possible: custom agents, retrieval over the codebase, and an evaluation loop that improves its own prompts.

How it works Close
  • A library of purpose-built agents for production incident investigation, database rightsizing, cost analysis, and documentation generation.
  • MCP servers exposing internal observability and codebase retrieval to agents as first-class tools.
  • An adversarial verification pattern — a second agent re-derives every number from source and labels each claim as measured, inferred, or assumed.
  • A scheduled retrospective loop that evaluates past agent runs against real outcomes and proposes concrete prompt improvements.
  • Claude
  • MCP
  • RAG
  • Python
Architecture Ongoing

Modernization and cost architecture

Ownership of a modernization roadmap spanning the database, infrastructure, and application layers of a large legacy estate.

How it works Close
  • A cloud cost program targeting a 90% reduction in data platform spend through workload consolidation, compute rightsizing, and scheduling redesign.
  • Database end-of-life migration strategy — options analysis and TCO comparison across managed engine choices.
  • Application server migration, scheduled-job decomposition to serverless, read replicas and connection pooling, caching expansion, and N+1 query elimination.
  • Rightsizing analysis across managed databases and compute fleets, grounded in measured utilization rather than defaults.
  • AWS
  • Aurora / MySQL
  • Terraform
  • Java
  • Spring Boot
02

Capabilities

AI & machine learning

  • Agentic system design
  • LLM application architecture
  • Retrieval-augmented generation
  • Evaluation harnesses
  • Computer vision pipelines
  • AI governance & guardrails

Cloud & platform

  • AWS solution architecture
  • Amazon Bedrock
  • Serverless & containers
  • Infrastructure as code
  • Cloud cost optimization
  • Observability

Engineering

  • Java 21 / Spring Boot
  • Distributed systems
  • Aurora / MySQL at scale
  • Legacy modernization
  • API & integration design
  • Capacity planning

Leadership

  • Technical leadership
  • Teams up to 20 engineers
  • Architecture decision records
  • TCO & business cases
  • Executive communication
  • Cross-functional delivery
03

About

I'm the Principal Software Architect behind a baggage tracing and lost-and-found platform used by major North American airlines. Sixteen-plus Java services, Aurora, a schema with two decades of accumulated meaning, and a lot of AWS.

For the last two years my focus has been applied AI in a real enterprise — not prototypes on clean data, but systems that have to survive a security review, a budget conversation, and an operations team that will notice immediately when something is wrong.

Before this I spent six years leading delivery teams of up to twenty engineers, building data and integration platforms for a major US airline. I moved to the architect track by choice, because I wanted to stay close to the build. Having now stood up an AI capability largely single-handed, I'm looking to do it again with a team behind it.

I sit between the domain and the technology. I write the architecture, and I write the one-pager that gets it funded.

04

Writing

05

Let's talk

I'm open to AI architecture and AI engineering leadership roles, and I'm always happy to compare notes with anyone building agentic systems against a legacy estate.

[email protected] LinkedIn