Skip to content
Open to AI infrastructure roles

Building scalable AI infrastructure and cloud-native systems

Hi! I'm Yash, a Software Development Engineer in Arizona building LLM infrastructure, evaluation systems and cloud platforms.

Show my work
Walnutech PBC

Vantion: 0 to 200+ users, ~90% of the implementation

AI-native engineering toolkit, shipped to 8 repositories

Daily drivers

My tech stack

  • AWS
  • Terraform
  • Docker
  • Python
  • TypeScript
  • Go

MS and BS in Computer Science from ASU, BS Summa Cum Laude

Current role

Cloud-native AI infrastructure at Walnutech PBC

Open to AI infrastructure roles

Connect on LinkedIn

Selected engineering work

pathwise: Trajectory Evaluation, cover image

pathwise: Trajectory Evaluation

Open-source (Apache-2.0) evaluation of how an agent reached an answer, not just the output.

check git repo
Vera: Agentic Personal Assistant, cover image

Vera: Agentic Personal Assistant

Production PWA with tool use, a read/write/act permission model, and per-call cost accounting.

private deployment
MCaaS: Model Compression as a Service, cover image

MCaaS: Model Compression as a Service

MS Thesis: Achieved 14x model size reduction and 3.4x faster inference with less than 4% accuracy loss.

check git repo
LLMOps: Serverless LLM Service, cover image

LLMOps: Serverless LLM Service

Serverless LLM service on AWS with Terraform IaC, Langfuse telemetry, and GitHub Actions OIDC CI/CD.

check git repo
Pulumi + Snowflake IaC Pipeline, cover image

Pulumi + Snowflake IaC Pipeline

Infrastructure-as-code pipeline provisioning AWS and Snowflake resources with Pulumi.

check git repo

My work experience

  1. Software Development Engineer @Walnutech PBC

    • Built Vantion, the company's product, from zero to 200+ users, personally responsible for ~90% of implementation across backend, ML, infrastructure, console and frontend
    • Replaced a 3GB MindsDB sidecar with a pgvector hybrid search engine over 43,000+ scholarships, cutting average query latency by 5x, eliminating 120s bootstrap time and reclaiming 4GB memory per ECS task
    • Built production LLM evaluation harnesses (LLM-as-judge with calibrated scoring and regrounding, multi-turn A/B, reranker calibration) and ran model-selection studies that drove production decisions on measured quality and cost rather than intuition
    • Authored an organization-wide AI-native engineering toolkit: 15 workflow skills, custom agents and safety hooks distributed to 8 repositories via an internal plugin marketplace
    • Designed cloud-native infrastructure across a multi-account AWS organization with IAM Identity Center and SCP guardrails, reducing access requests by 65%

  2. Software Developer @United Portraits

    • Built SMS Manager, a Twilio messaging platform with batch scheduling, opt-out tracking and quiet-hours enforcement
    • Raised the organization's Twilio compliance score from 78% to 94% within 4 weeks, sustaining 98.6% weekly compliance
    • Shipped on Angular 21 and Spring Boot 4 with Keycloak, PostgreSQL and Docker Compose behind Nginx

  3. Cloud & ML Researcher @VISA Research Lab

    • Engineered a serverless infrastructure automation pipeline (MCaaS) on AWS scaling to 100+ executions/month, reducing manual ML compression effort by 90% and deployment failures by 42%
    • Orchestrated distributed workflows using AWS Step Functions, achieving 14x model size reduction and 3.4x faster inference with less than 4% accuracy loss
    • Built on Python, PyTorch, Docker and Terraform across ECS Fargate, S3, Lambda and Step Functions

My approach

  1. Phase 1

    Measure before optimizing

    Turn subjective output quality into reproducible signal: LLM-as-judge with calibrated scoring and regrounding, multi-turn A/B evaluation, and model-selection studies decided on measured quality and cost rather than intuition.

  2. Phase 2

    Build it reproducibly

    Infrastructure as code with Terraform and Pulumi, containers on ECS Fargate, and telemetry through Langfuse and CloudWatch. CI/CD runs on GitHub Actions with OIDC, per-task IAM roles and secrets held in SSM.

  3. Phase 3

    Enforce the rules mechanically

    Operational rules belong in hooks, not documentation. Pre-execution checks refuse a production deploy from the CLI and refuse a database write against the wrong project, each one added after that failure mode happened once.