Dmitry Starodubtsev

Backend & AI Engineer

Dmitry
Starodubtsev

Building production RAG pipelines, tool-calling agents with memory, and multi-tenant backends with scalable AWS infrastructure. Michigan State CS '26.

3,200+ daily users

7th-largest K-12 district in Illinois

95%+ retrieval accuracy

5 university partners

GitHubLinkedInstarodu5@gmail.com

About

Backend and AI engineer specializing in production RAG pipelines, tool-calling agents with memory, and multi-tenant backends with scalable AWS infrastructure. At APS Data Technologies I've shipped systems supporting 3,200+ daily users at the 7th-largest Illinois K-12 district and production RAG for 5 university partners.

I build things that hold up in production, not just in demos.

Career

Experience

August 2026 — Present

Applied AI Engineer — Copilot & Excel Team

Microsoft — Part-Time Contract (Remote)

Design LLM evaluations for Copilot and Excel workflows, building test cases and scoring rubrics for multi-step reasoning, code generation, and tool use while identifying systematic failure modes and expanding adversarial coverage.

April 2025 — August 2026

Software Developer — AI & Backend

APS Data Technologies — EdTech B2B SaaS

Built a stateful FastAPI agent with tool calling, per-user memory, and RBAC-aware execution across enterprise APIs, cutting routine ERP task time 5×. Architected containerized services across AWS and GCP, sustained 5K+ RPS in load tests, and shipped durable async screening, multi-tenant Django RBAC, and production RAG.

September 2024 — April 2025

Software Engineer

Safety Straw — Seed-stage Consumer Safety Startup

Built the commerce backend with checkout APIs, payment-state reconciliation, and idempotent asynchronous order workflows. Shipped the React/Next.js storefront and automated CI/CD, cutting deployments from 30–45 minutes to 5–8 minutes while improving mobile load time by 23%.

September 2024 — May 2025

Operations & Growth Lead

Imagine Software — Student Organization

Scaled student tech org 55% from 450 to 700 members. Conducted 20+ technical interviews for client project staffing.

August 2023 — August 2024

Data Automation Engineer

Michigan State University

Cut archival-data processing time 10× across thousands of newspaper records with a Python ETL pipeline that normalized metadata, detected missing records, and automated archive lookups. Recovered missing publication records and standardized date and metadata formats for downstream research.

Aug 2022 — May 2026

B.S. Computer Science, Business Minor

Michigan State University

GPA 3.75. SpartaHack XI Blockchain Track 3rd Place (among 100+ teams).

Case Studies

Selected Work

Production Work

ERP Platform for K-12 Schools

2025 — Present

Multi-tenant ERP connecting HR, payroll, background checks, and staff workflows for K-12 school districts. Supports 3,200+ daily active users at the 7th-largest K-12 district in Illinois and processes 500 AI-processed resumes per day. The AWS stack sustained 5K+ RPS in k6 load tests.

Software Engineer @ APS Data Technologies

Django RESTCeleryNext.js 15PostgreSQLAWSTerraform

Problem

  • Candidate screening ran sequentially and routine HR and payroll work spanned disconnected enterprise systems, while each district needed configurable authorization without custom permission code.

What I Built

  • Async candidate-screening pipeline (Django + Celery + Redis) for document parsing, AI scoring, and concurrent evaluation, with idempotent retries and row-level locking.
  • Runtime-configurable RBAC architecture that lets districts define roles, permissions, and hierarchical resource scopes without per-district authorization code.
  • Tool-calling LLM agent with per-user memory and integrations across enterprise HR, payroll, and background-check systems.

Impact

  • Candidate screening cut 10× from 2 hours to 12 minutes and routine HR and payroll tasks cut 5×. The AWS stack supports 3,200+ DAU and 500 AI-processed resumes per day, and sustained 5K+ RPS in k6 load tests.
Demo

Admissions & Career Platform for University Partners

2025 — Present

Production RAG platform serving 5 university partners. Reaches 95%+ retrieval accuracy across a 1K+ document corpus using async ingestion, OpenAI embeddings, Qdrant semantic retrieval, cross-encoder reranking, and evaluation across 100+ test cases.

Software Engineer @ APS Data Technologies

Django RESTQdrantOpenAICross-encoderAWS S3/ECS/RDSDocker

Problem

  • University partners needed accurate retrieval over a growing document corpus, with a repeatable way to measure whether retrieval changes improved results instead of relying on anecdotal responses.

What I Built

  • Async document-ingestion pipeline that creates OpenAI embeddings and indexes content in Qdrant for semantic retrieval.
  • Two-stage RAG pipeline combining vector retrieval with cross-encoder reranking to improve precision after the initial semantic search.
  • Retrieval evaluation suite covering 100+ test cases against a 1K+ document corpus to measure accuracy and validate pipeline changes.

Impact

  • Deployed with 5 university partners and achieved 95%+ retrieval accuracy across the evaluated document corpus.
Demo

Safety Straw

Sep. 2024 — Apr. 2025

E-commerce and customer-support systems for a seed-stage consumer safety startup. Built a Node.js/Express backend with Stripe, asynchronous order workflows, and an AI support agent grounded in product, order, and shipment context.

Software Engineer @ Safety Straw (Seed-stage Consumer Safety Startup)

Node.jsExpressStripeREST APIsDockerGitHub Actions

Problem

  • Inbound customer-support inquiries had an initial response time of 4–12 hours, while the e-commerce backend needed reliable payment, order, and notification workflows.

What I Built

  • Node.js/Express e-commerce backend with Stripe checkout, payment-state handling, order lifecycle APIs, and asynchronous notification and order-processing workflows.
  • AI support agent grounded in product, order, and shipment context, with escalations pre-filtered for manual support.
  • Docker and GitHub Actions CI/CD pipeline covering testing, migrations, staging builds, and zero-downtime deployment.

Impact

  • Automated 80%+ of inbound support inquiries and cut initial response time from 4–12 hours to under 30 seconds. Reduced deployments from 30–45 minutes to 5–8 minutes.

Other Projects

Multi-Agent Simulation Engine

2026

Real-time civilization simulation supporting up to 50,000 concurrent agents. The scheduler is world-agnostic: any simulation plugs in by implementing a single DoAction interface, and the scheduler stays completely blind to agent and action types. Scaled using a 4-tier priority system (critical/gameplay/environment/misc) with async job queues, separating deterministic combat from construction tasks to prevent frame stalls.

Product Lead — Team of 30

Scheduling Architecture

  • Implemented stride scheduling where processes track virtual time position, with stride inversely proportional to priority. The scheduler always advances the furthest-behind process, ensuring fair proportional CPU allocation across 50,000+ agents.
  • Organized execution into 4 tiers (CRITICAL 40%, GAMEPLAY 30%, ENVIRONMENT 20%, MISC 10%) with independent Scheduler instances per tier via composition.
  • Set per-tier budgets with soft enforcement for CRITICAL (250ms overage allowed) and hard cutoff for others, preventing lower-tier processes from starving critical gameplay loops.
  • Built in C++23 compiled to WebAssembly via Emscripten, designing clean interfaces between world state, agent behaviors, and rendering systems for a distributed 30-person team.
  • Verified correctness across unit tests (Scheduler add/remove/priority/peek), integration tests (budget enforcement, tier isolation, edge cases), and starvation scenarios.
C++23WebAssemblyEmscriptenStride SchedulingUnit Testing
GitHub

OpsCore

2026

Demo showing three production patterns: capability-based RBAC (why deny-by-default scales better), two-stage RAG (embedding + cross-encoder tradeoff), and agentic systems with tool execution and memory.

Demo Project — open source

  • RBAC: Districts need self-service role management — capability-based, deny-by-default, so a missing grant means denied without maintaining deny rules.
  • RAG: Support agent needs 95%+ accuracy at sustainable cost — two-stage pipeline uses embedding for recall, cross-encoder for precision.
  • Agent: Resume screening must complete even if client disconnects — Celery task, not streaming HTTP.
Django RESTNext.js 14PostgreSQLRedisCeleryQdrantClaude API
GitHubREADME

QuiKard

2025

Built a digital business card service: create in under a minute, export to Apple Wallet without signing up, share via NFC tap.

Solo Developer

  • Next.js 15 + TypeScript frontend with Tailwind styling, FastAPI + SQLAlchemy backend on PostgreSQL/SQLite.
  • Integrated Apple Wallet Pass API to generate .pkpass files on-the-fly, enabling NFC tap sharing directly from iPhone Wallet.
  • Each card gets a unique shareable URL and auto-generated QR code for distributing physical cards without typing or links.
Next.js 15FastAPIPostgreSQLApple Wallet APITypeScriptDocker
GitHub

CourseChecker

2025

Course review aggregator where only verified university students (via school email) can submit reviews. Tracks 5 rating dimensions (overall, difficulty, materials, workload, fairness) per professor and semester.

Lead Developer (85 of 153 commits)

  • Tracks 5 independent rating axes (overall, difficulty, materials, workload, fairness) with granularity per professor and semester.
  • Implemented pagination and incremental data fetching so the UI stays responsive as review count grows, avoiding full dataset loads on mount.
  • Used Supabase RLS to enforce per-student data isolation, handle school email domain validation, and provide OAuth without custom auth infrastructure.
Next.js 14SupabaseGPT-4TypeScriptTailwind
GitHub

Volunteer Matchmaker

2026

Volunteer matching system using composite scoring across proximity, skills, urgency, and reliability. Integrated Ethereum smart contracts for on-chain donation tracking. Won SpartaHack XI Blockchain Track 3rd Place.

Team of 4 — SpartaHack XI | Blockchain Track 3rd Place

  • Scoring algorithm weights proximity (geodistance), skill overlap, job urgency (time-decay for stale requests), and volunteer reliability rating to produce ranked matches.
  • Built real-time messaging between matched volunteer and requester without exposing personal contact info, using JWT-authenticated channels.
  • Deployed VolunteerLeaderboard smart contract on Ethereum for transparent donation tracking, integrated to frontend via Ethers.js.
Next.js 15Django RESTGemini AISolidityZustand
GitHub

Toolkit

Technologies

AI / ML

  • RAG Systems
  • Vector Search (Qdrant)
  • LLM Integration (GPT-4, OpenAI API)
  • Embeddings
  • Cross-encoder Reranking
  • Agentic Systems (Tool Calling, Memory)
  • Retrieval Evaluation

Backend / Systems

  • Python
  • Django REST
  • FastAPI
  • Celery
  • Node.js / Express
  • REST APIs
  • Async Processing
  • RBAC
  • Stripe
  • C++
  • SQL

Frontend

  • TypeScript
  • React
  • Next.js
  • HTML/CSS
  • Tailwind CSS
  • Zustand

Infrastructure

  • AWS (ECS, RDS, S3)
  • Terraform
  • Docker
  • CI/CD (GitHub Actions)
  • CloudWatch
  • k6
  • Autoscaling

Databases

  • PostgreSQL
  • Redis
  • MongoDB
  • Qdrant
  • SQLite