_{The Claude Code figures in the header are regenerated nightly from a public gist whose revision history is the audit trail.}

I build AI infrastructure and the evaluation that tells you whether it works.

Right now that means Vantion, an AI product I took from zero to 200+ users and wrote ~90% of across backend, ML, infrastructure, console and frontend. For most of that time I was the only engineer at a company of three. That is the denominator the number needs.

Sep 2025 to now · full-time since April 2026

• Built Vantion from zero to 200+ users and wrote ~90% of the implementation across backend, ML, infrastructure, console and frontend. Only engineer at a three-person company for most of that time. Two interns have since joined.
• Built LLM evaluation harnesses that run in production: LLM-as-judge with calibrated scoring and regrounding, multi-turn A/B evaluation, a reranker calibration harness. One run takes 20 randomly selected student profiles and produces 800 judgments, which turns output quality into a measurement instead of an opinion.
• Ran model-selection studies that decided production, including Haiku over Sonnet on a measured quality and cost tradeoff rather than on intuition.
• Authored the organization's AI-native engineering toolkit: 15 workflow skills, custom agents and safety hooks distributed to 8 repositories through an internal plugin marketplace, with every repository audited on a 0 to 5 agent-readiness rubric and the rollout briefed to leadership.
• Designed a multi-account AWS organization (dev, test, prod) with IAM Identity Center, SCP guardrails and environment-specific roles. Access requests fell 65%.
• Replaced a 3GB MindsDB sidecar with a pgvector hybrid search engine over 43,000+ scholarships: average query latency down 5x, 120s of bootstrap time gone, 4GB reclaimed per ECS task. The speed came from deleting the network hop and the boot, not from tuning the search.

_{Python, FastAPI, LangChain, LangGraph, Langfuse, pgvector, MCP, Anthropic and OpenAI APIs, Docker, AWS (ECS Fargate, ALB, CloudWatch, IAM Identity Center, Secrets Manager), Terraform, GitHub Actions with OIDC.}

Jan 2026 to Mar 2026

• Built SMS Manager, a Twilio messaging platform with batch scheduling, opt-out tracking and quiet-hours enforcement. Twilio compliance went from 78% to 94% in four weeks and held at 98.6% weekly.
• Two months, and I left for AI work specifically, evaluation and model behaviour. The dates overlap Walnutech because they were different commitments: full-time here, a few hours a week there while Walnutech was still an internship.

_{Angular 21, TypeScript, Fabric.js, Three.js, RxJS, Spring Boot 4, Java 21, Twilio, Keycloak, PostgreSQL, Docker Compose, Nginx.}

Jan 2024 to Aug 2025

• Built MCaaS, a serverless model-compression pipeline on AWS that scaled to 100+ executions per month, cut manual compression effort by 90% and deployment failures by 42%.
• Orchestrated distributed workflows on AWS Step Functions: 14x model size reduction and 3.4x faster inference at under 4% accuracy loss.

_{Python, PyTorch, Bash, AWS (ECS Fargate, S3, Lambda, Step Functions), Docker, Terraform. Also my master's thesis.}

Review means using the thing. Every pull request on my team runs automated tests plus a Claude and a Codex review. Bugs still reached production, because reviewers were approving on the strength of reading a diff. Each pull request now raises an ephemeral environment, a container with the database and the frontend, and posts the link into the conversation with a guideline for what to check. The reviewer opens it before approving, and the container is destroyed on merge. Fewer bugs reach production.

A confident model answer is a starting point, not a conclusion. On MCaaS I implemented unstructured pruning because an LLM said it would reduce model size. It zeroes weights, it does not remove them, so the file on disk was exactly the same size. I had already taken that result to my professor.

Ask a model for structure, not for the answer. Matching student profiles against scholarship eligibility with an LLM directly was bad. The match should stay deterministic, SQL and explicit rules, with the model converting eligibility prose into machine-checkable rules, ontology fields and entity relationships. Structure can be verified. An answer has to be trusted.

Measured, not claimed. yashshelar.com ships behind its own npm run verify gate, and its animation performance is graded by an outside tool rather than described in prose: , re-measured on every load.

• MS Computer Science, Arizona State University, Aug 2025. Thesis: Model Compression as a Service for Resource-Constrained Edge Devices.
• BS Computer Science, Arizona State University, Dec 2022. Summa Cum Laude, New American Scholar, Dean's List.

From: 08 August 2025 - To: 31 October 2026

Total Time: 385 hrs 49 mins

Markdown                   115 hrs 36 mins       >>>>>>-------------------   24.91 %
TypeScript                 81 hrs 10 mins        >>>>---------------------   17.49 %
Other                      78 hrs 16 mins        >>>>---------------------   16.87 %
Python                     74 hrs 16 mins        >>>>---------------------   16.01 %
JavaScript                 24 hrs 28 mins        >------------------------   05.27 %
YAML                       21 hrs 23 mins        >------------------------   04.61 %
Terraform                  11 hrs 33 mins        >------------------------   02.49 %
SQL                        10 hrs 16 mins        >------------------------   02.21 %
JSON                       9 hrs 10 mins         -------------------------   01.98 %
Bash                       8 hrs 43 mins         -------------------------   01.88 %

LinkedIn is the way to reach me.