Marian Dumitrascu
I build production GenAI systems on AWS — designed for the teams that own them.
Principal Cloud Solution Architect
AWS · Generative AI · Multi-Tenant SaaS
📍 Houston, TX • 20+ years experience
Technical Expertise
Comprehensive skills across AI/ML, Cloud Infrastructure, and Data Analytics
Amazon Bedrock & Agentic AI
expertRAG & Knowledge Bases
expertMulti-Agent Orchestration
expertVision LLMs
expertMulti-Provider LLM Patterns
expertAWS AI/ML Services
expertAI-Assisted Delivery
expertQuantum Computing
advancedAWS Cloud Architecture
expertAWS Compute & Containers
expertAWS Data & Analytics
expertAWS Storage & Networking
expertMulti-Tenant SaaS
expertStripe Billing Integration
expertInfrastructure as Code
expertDevOps & Observability
expertLanguages
expertWeb Development
expertData Engineering
expertData Science & ML
expertDatabases
expertEngineering Practice
expertFeatured Projects
Real-world solutions demonstrating expertise in AI/ML, cloud architecture, and data science
Nutrition-assessment, recipes, meal-planning, grocery-list, and gamification platform for a health and wellness startup, integrated with the Z-Code Social Determinants of Health survey and a closed feedback loop driving personalized recommendations. HIPAA-aware design.
Migrated AI services from direct Bedrock API calls to deployed AgentCore Runtime agents (Assessment, Grocery, Chat, Meal Plan plus dietary and safety validators). Centralized prompt-template management in Admin UI; backend retrieves and compiles templates with per-user context. Recipes & Meal Plans powered by Bedrock Knowledge Bases with healing-herbs retrieval. Six admin-panel modules plus CSV import/export and backup/restore with dry-run. 320+ tests passing across backend, frontend, mobile, and E2E suites.
POC for an industrial QA partner: automated detection of gas leaks during underwater pressure-testing of meters from an 8-camera rig generating 2.2 TB of 4K video across 794 files. Interactive dashboard with PASS/FAIL determinations, frame-level evidence viewer, AI reasoning per call, and CSV/Excel exports.
Strategic pivot from YOLO object detection to Bedrock Claude vision LLM cut POC timeline from 3–4 weeks to 4 days and initial cost from $5K+ in annotation labor to under $50 in API spend, while gaining natural-language explainability per detection. v5 bubble-classification breakthrough distinguishing condensation/brushing artifacts from true gas-leak indicators: 100% API success across 783 frames analyzed; zero severe false alarms; 79.8% correct false-positive detection on brushing activity. Multi-color-space CV pipeline narrowed search space from 6+ hours of footage to 12.3 minutes (94% reduction).
Production-ready platform for natural-language venue search, AI-driven recommendations, AI receipt processing, and Stripe metered commission billing. Delivered in a 2–3 week intensive sprint with 362 tests at 100% backend coverage; shipped 7 versions for incremental client feedback.
Hybrid semantic + structured + full-text search using Titan V2 embeddings (1024-dim) in pgvector with HNSW indexing combined with PostgreSQL tsvector and per-user personalization signals; sub-500 ms p95 search latency. Dual-provider AI architecture (AWS Bedrock primary, OpenAI fallback) with per-call cost tracking, multi-level caching, and request deduplication; ~50–70% cost reduction vs single-provider baseline; 99.9% effective uptime. AI receipt OCR via Bedrock Claude Sonnet 4 vision API at 85%+ accuracy with fuzzy venue matching driving Stripe metered billing.
Multi-sided learning marketplace with educator and consumer roles, AI-assisted course creation with Mux video processing, RAG-powered course search, and Stripe Connect educator payouts.
Bedrock Knowledge Bases + RAG over course content with reranking and SSE-streamed UI for live progress visibility. AI-assisted course-creation pipeline integrating Mux video webhooks with FastAPI orchestration; replaced originally-planned Step Functions design with simpler in-process orchestration. Architecture rewritten mid-project from server-actions to explicit FastAPI / Next.js separation with documented API contract; full E2E test surface with Playwright + Vitest + Pytest.
Production C++ AWS Lambda backend (43 endpoints) plus React 19 / Vite frontend automating Terraform-driven AWS provisioning with real-time SSE monitoring and broadcast media-services integration for a media-services company.
Deliberate two-phase delivery: 4-week Python / FastAPI prototype to validate requirements (28 MSL tests + 17 integration tests), then 3-week C++ migration to meet client tech preference; achieved sub-50 ms response times (vs 100–150 ms in Python) and lower Lambda execution cost. 100% API compatibility, full test suite ported.
Open-source methodology and document templates for managing complex software projects with AI coding assistants — splitting strategic planning, task delegation, and execution across role-specific chat sessions, backed by living-document artifacts (CURRENT-STATUS.md, DELEGATION-TRACKER.md, THINKING-LOG.md, WORKER-PROMPTS/).
Used in production across the wellness platform's 320+-test buildout, the educational marketplace's 12-session migration, and other engagements. Public repository plus an in-depth LinkedIn writeup. The working method behind every recent client engagement.
GitHub Activity
ML Demo
Try this interactive sentiment analysis model. Enter any text and see how AI interprets emotional tone in real-time.
Try these examples:
About Me
Principal Cloud Solution Architect with 20+ years across software architecture, AI/ML, and data science — focused for the last four years on shipping production-grade GenAI systems on AWS. Strength: turning ambiguous business problems into running architectures, from discovery and scoping through reference implementation and clean handoff. Deep current practice in agentic AI on Amazon Bedrock (Claude family, AgentCore Runtime, Strands Agents SDK, Knowledge Bases / RAG), multi-tenant SaaS platforms, and AI-assisted delivery methodology.
Get In Touch
Interested in collaborating, hiring for a Principal-level GenAI role, or just want to chat about an architecture problem? Send a message and I'll get back to you within a business day.