Skip to content

Yash Shelar, AI infrastructure engineer

I'm Yash, building LLM infrastructure, evaluation systems and cloud platforms in Arizona.

rednote-gate checked against the live site on

Dev, Test and Prod in one AWS organization

Dev, Test and Prod live in separate accounts under IAM Identity Center. Service control policies set the guardrails, and each environment has its own roles.

  • 65% fewer access requests
IAM Identity CenterDevTestProdSCP guardrails

A 3 GB sidecar, replaced by pgvector

Hybrid search over 43,000+ scholarship records now runs on pgvector, where a 3 GB MindsDB sidecar used to sit.

  • 5x lower average query latency
  • 120 s bootstrap removed
  • 4 GB memory back per ECS task
3 GB sidecarpgvector43,000+ records

Model compression on Step Functions

A serverless pipeline, orchestrated with AWS Step Functions, compresses models and deploys them. It runs 100+ times a month.

  • 14x smaller models
  • 3.4x faster inference
  • <4% accuracy loss
  • 90% less manual effort
  • 42% fewer failed deploys
routeprunedistillquantizeevaluateStep Functions states

Evaluation harnesses that pick the model

LLM-as-judge with calibrated scoring, multi-turn A/B evaluation and a reranker calibration harness make output quality reproducible. One study chose Haiku over Sonnet on measured quality and cost.

  • Haiku chosen over Sonnet
judgeHaikuSonnet

Open source

Figures and quotes are from each project's README, as it read on 2026-10-08.

  1. convene

    github.com/YashShelar007/convene

    Local inference layer for Python on top of a logged-in Claude Code.

    9.5x cheaper warm: $0.00281 against $0.02670.

    ~10x throughput, 0 rate-limit errors at a burst of 20.

    Samples are small and each one says so. Tables and raw figures are in FINDINGS.md.

    Python 3.11+, zero dependencies

  2. claude-workflows

    github.com/YashShelar007/claude-workflows

    Two Claude Code plugins: blind-review, where two model families review a PR blind and a third rules where they disagree, and site-preflight, a twenty-item launch checklist run against a live site.

    blind-review: two reviewers from different model families read a PR's diff and brief and nothing else, in parallel, with byte-identical prompts.

    Where they disagree, a third model family adjudicates.

    Claude Code plugins, Node 22 or newer

  3. rednote-gate

    github.com/YashShelar007/rednote-gate

    MCP server for one throwaway RedNote account.

    5 read tools run straight away. 6 write tools only queue. 4 local tools never touch RedNote.

    Every write waits in a queue until a human clicks Approve on a local web page.

    It works by browser automation, which is against RedNote's terms of service.

    TypeScript, Playwright, Node 20 or newer

  4. claude-desktop-session-restore

    github.com/YashShelar007/claude-desktop-session-restore

    Rebuilds the Claude Desktop session list from CLI transcripts.

    The 69 macOS records carry 44 distinct fields in 44 distinct field-set signatures, from 19 to 33 fields per record.

    Python 3.9+, PowerShell 5.1+, no dependencies

Ask Friday about my work

Friday answers from my résumé and nothing else, and the section behind each answer opens under it.

Work and education

WhenWhat
Sep 2025 to present

Software Development Engineer

Walnutech PBC

Converted to full-time April 2026.

  • Built Vantion, the company's product, from zero to 200+ users; personally responsible for ~90% of implementation across backend, ML, infrastructure, console and frontend.
  • 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; audited every repo on a 0 to 5 agent-readiness rubric and briefed leadership on the rollout.
  • Designed and implemented cloud-native infrastructure across a multi-account AWS organization (Dev/Test/Prod) with IAM Identity Center, SCP guardrails, and environment-specific roles, reducing access requests by 65%.
Jan 2026 to Mar 2026

Software Developer

United Portraits

  • Built SMS Manager, a Twilio messaging platform with batch scheduling, opt-out tracking, and quiet-hours enforcement, raising the organization's Twilio compliance score from 78% to 94% within 4 weeks and 98.6% weekly compliance.
Jan 2024 to Aug 2025

Cloud & ML Researcher

VISA Research Lab

  • Engineered a serverless infrastructure automation pipeline (MCaaS) on AWS that scaled 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.
Aug 2025

MS Computer Science

Arizona State University

Dec 2022

BS Computer Science

Arizona State University

Summa Cum Laude