I build AI systems where agents have roles, memory, and authority — not just prompts.

Founder @ vctrx.dev[ Active Build ]

Kirtan Thummar

AI Orchestration & Infrastructure Engineer

# About />

I've spent the last few years trying to answer one question: what does it actually look like when AI works the way a real team does?

# Experience & Education />

7

M.S. Information Technology — Arizona State University

Aug 2024 – May 2026

GPA: 4.00/4.00 — Focus on AI orchestration, cloud infrastructure, and developer tooling.

6

Full Stack AI Engineer — Opportunity Hack

Oct 2025 – Jun 2026

Turned a hackathon win into a live product. Rebuilt nmtsa.org from scratch with a team of 3 — full CMS, admin panel, payment integration, and AI features. The site serves a real nonprofit community.

5

Software Engineer Intern — ASU Learning Engineering Institute

Jan 2026 – May 2026

Sole engineer on the Critical Code Reviewer project — prompt architecture, diff parsing, patch-mapping logic, and inline comment delivery via GitHub Actions. Designed a self-improving eval pipeline with LLM-native quality metrics.

4

Founder & Software Engineer — Vctrx

Nov 2024 – Present

Building an AI orchestration platform for role-based, spec-driven software development. Sole architect and engineer. Agents map to real org roles with persistent memory and scoped authority.

3

Software Engineer (Cloud & AI Systems) — Braincuber

Feb 2024 – Aug 2024

Led a team of 4 building an AI for Health product end-to-end. Owned AWS infrastructure, OpenAI API integration, backend services, and parts of the frontend. Made architectural decisions, not just executed tickets.

2

Junior Software Engineer — Braincuber

Sep 2022 – Jan 2024

Full-stack engineering with React, Next.js, backend APIs, and cloud infrastructure on AWS and GCP. Grew into solution architecture work beyond the original scope of the role.

1

B.E. Information Technology — Gujarat Technological University

Oct 2019 – Apr 2023

Built the engineering foundation that everything else sits on top of.

# Projects I've shipped />

Vctrx — AI Orchestration Platform

Founder & Solo Engineer| Active build| vctrx.dev

Multi-agent orchestration platform where a PM agent reviews requirements, a developer agent writes code to spec, a security agent audits the output, and a DB agent handles schema — each with their own memory, permissions, and defined scope of authority. Persistent memory layer backed by RAG, a vector database, and PostgreSQL, so agents have retrievable context across long-running workflows, not just the last N tokens. Inter-agent communication is MCP-compatible. The cloud module accepts natural language and generates Terraform change sets via Amazon Bedrock, so a non-engineer can provision infrastructure without touching a config file. The whole bet is that structure beats intelligence at scale.

Paperlyft — Government Accessibility SaaS

1st Place, AI + Elections Hackathon (ASU, Feb 2026)| Active build

Government PDFs are inaccessible to anyone using a screen reader or assistive technology. The ADA Title II amendments require public entities to meet WCAG 2.1 AA by 2026–2027, and most have no tooling to comply. Paperlyft's pipeline takes a government PDF and outputs verified WCAG 2.1 AA-compliant HTML. I mapped 36 actionable WCAG criteria across all four principles into the prompt pipeline so the model has precise compliance targets, not vague guidelines. Verification runs through PAC 3 and axe-core. The human review flagging system routes edge cases the pipeline can't auto-resolve instead of guessing wrong and shipping something non-compliant — AI that knows what it doesn't know is harder to build than it sounds.

Critical Code Reviewer (CCR) — ASU Learning Engineering Institute

Sole Engineer| Jan 2026 – May 2026

A GitHub Actions workflow that reviews pull requests using an LLM backend. Analyzes diffs, applies a structured multi-layer evaluation rubric, and posts inline comments anchored to specific lines — not at the bottom of the PR where nobody reads them. The hard part was patch-mapping logic: figuring out which line in the current file maps to which position in the unified diff, solved with fuzzy text search against unified diffs. After the core system shipped I voluntarily designed a self-improving eval pipeline with LLM-native metrics for accuracy, relevance, and false-positive rate — feedback loops that let the system improve its review quality without manual tuning. Most engineers ship the thing and move on. I shipped the thing, then built the thing that measures it.

NMTSA.org — Full Nonprofit Rebuild

Winner, Opportunity Hacks 2025| Live in production| nmtsa.org

Neurologic Music Therapy Services of Arizona needed more than a website. Broken contact forms, no CMS, no payment processing, no way to update content without a developer. We won the hackathon in a weekend, then kept building for 9 months. Shipped a full rebuild from scratch with a team of 3 — Next.js frontend, Node.js backend, PostgreSQL, Stripe payments, a full CMS, admin panel, and AI features integrated throughout. It's live with real users: therapists and patients who depend on it daily. Most hackathon projects die on Monday. This one didn't.

DeviceLoop — E-Waste & Digital Access

Winner, SunHacks 2025| 600+ participants

Companies retire laptops after 3–4 years — not because the hardware's broken, but because data privacy policies make device reuse feel legally risky. The devices get shredded while first-gen students try to learn to code on phones. DeviceLoop is a certified, auditable data-wiping pipeline that meets enterprise security standards, paired with a platform that routes verified devices to first-gen students and community organizations. One solution for a data privacy problem and a hardware access problem at the same time. Won among 600+ participants at SunHacks 2025 — the idea was strong enough that we kept getting asked about it after the event.

# Skills & Expertise />

# Awards & Recognition />

1st Place — AI + Elections Hackathon

ASU (Feb 2026)

Built the foundation for Paperlyft — a civic AI tool co-designed with election officials to strengthen accessibility, transparency, and public trust in elections. Recognized under ASU's Principled Innovation framework.

Winner — Opportunity Hacks 2025

Oct 2025

Built a real-world tech solution for NMTSA over a weekend, then continued development for 9 months into a live production site serving a real nonprofit community.

Winner — SunHacks 2025

Sep 2025 | 600+ participants

Built DeviceLoop: a certified data-wiping and device donation platform connecting corporate device retirement to first-gen students, reducing e-waste in the process.

# Certifications />

AWS Certified Cloud Practitioner

Foundational AWS knowledge — cloud concepts, security, pricing, and services.

AWS Academy Graduate — Cloud Security Foundations

Security-specific credential covering IAM, encryption, compliance, and incident response on AWS.

+ 8 additional certifications on LinkedIn

# Volunteering />

Mathematics & Programming Tutor

Sep 2019 – Dec 2023

BAPS Swaminarayan Sanstha

Tutored students in mathematics and programming fundamentals for over 4 years while completing an engineering degree. Teaching fundamentals to others deepened my own understanding of them.

# Technical Philosophy />

Structure beats intelligence

A well-scoped agent with clear context will outperform a smarter model with no guardrails almost every time. I've seen this enough that it's not a hypothesis anymore.

Ship it, then improve it

Build the thing. See how it breaks in the real world. Then build the thing that measures it. You can't design good metrics in a vacuum — the system has to fail first.

AI as multiplier, not replacement

The right question isn't "can AI do this job?" It's "which parts of this job shouldn't need a person?" Get that right and you free people up for the work that actually requires judgment.

Accountability is a feature

If you can't tell me which agent made which decision and why, the system isn't production-ready. Auditability isn't an enterprise checkbox — it's how you debug, improve, and actually trust something over time.

Consistency over capability

I'd rather have an agent that does one thing reliably than one that does ten things unpredictably. Most real-world AI failures are consistency failures, not capability failures.

Context is architecture

Most people treat context as an afterthought — something you dump into a prompt and hope for the best. The way you scope it, structure it, and deliver it to each agent is the system design. Get context wrong and it doesn't matter how good the model is.

# What I'm Looking For />

I want to work on real AI systems — not internal demos, not pilot programs that never make it to production, not AI features bolted onto a product that doesn't need them.

Roles

AI Engineering (agents, orchestration, LLM pipelines), AI Infrastructure (model serving, eval systems), Full-Stack with heavy AI integration.

Location

Open to on-site, hybrid, or remote. Willing to relocate.

Environment

Hard AI problems, engineers trusted to make architectural decisions, reliability treated as a feature.