Prototype work with real infrastructure constraints.
Volico Data Centers — Oct 2025Internship
Software Developer Intern
Deerfield Beach, FL, USA
Low-latency AI voice assistant pipeline
PythonFlask APILightRAGTwilioLinux
Worked with a software engineer on a call-center voice assistant prototype; independently built the LightRAG retrieval layer for company/product knowledge.
Projects
Standalone systems that can keep evolving.
Open SourceProject
Prism
AI website customization extension
AI browser customization for live websites
Chrome ExtensionPython BackendAI UX
Customizes websites users already visit with anchored DOM context and patch-based CSS/DOM edits.
ActiveAI System
Clavis
Private portfolio operations layer
Telegram-controlled assistant for portfolio operations
Hermes AgentTelegramFastAPI
Private assistant workflow for farcas.dev maintenance, scheduled tasks, verification, and public-safe visitor Q&A.
ActiveArchive
AI Research Archive
Source-grounded AI catalogue
Source-grounded archive for AI research
JSON/JSONLStatic PublishingMetadata
Public archive for AI papers, models, datasets, tools, and repositories with normalized records.
Competitions
Compact challenge context, expandable on demand.
OpenAI Parameter Golf — 2026Competition
ETD Hybrid Architecture
OpenAI Parameter Golf competition
Compact transformer under a 16MB limit
ETDInt5 GPTQ1.1169 BPBVisualizer
OpenAI Parameter Golf submission with a compact ETD transformer, skip connections, quantization, and a companion visualizer.
May 2026IBM BOB
IBM BOB Hackathon
Hackathon prototype that became Prism
48 HoursAI Build Sprint
Original browser-customization prototype; current continuation lives under Prism.
2024HBN²
HBN² AML Challenge
AML fraud-detection machine-learning team challenge
Machine LearningTransaction Data
Team challenge around fraud and money-laundering detection from transaction datasets.
Technical Foundation & Education
The practical stack behind the work.
Detailed skill context is still available, but the first view stays compact.
CoreAPIs
Software Engineering & APIs
Backend APIs, auth, and integrations
RESTAuthDatabases
REST API design, integrations, auth, backend services, and database work.
AgentsAI
AI & Automation
AI agents and automation workflows
LLMsRAGVoice AI
LLM integration, agents, RAG, LightRAG, voice AI, and workflow automation.
SecurityTesting
Security & Networking
Security testing and traffic analysis
KaliBurpWireshark
OWASP-based testing, reconnaissance, packet inspection, and vulnerability assessment.
OpsLinux
Linux & Infrastructure
Self-hosted Linux infrastructure and deployment
DebianNginxProxmox
Linux administration, Raspberry Pi deployment, reverse proxies, SSL, and server workflows.
B.Sc. Computer Science plus mathematics and computer science high-school foundation.
Professional Experience · Prototype Foundation
Matei worked as a Software Developer Intern at Volico Data Centers. He worked with a software engineer to design and build a low-latency AI voice assistant prototype for call-center/customer-service use cases.
Researched open-source STT, LLM, and TTS model options against quality, latency, deployment, and integration constraints.
Proposed architecture approaches focused on reducing end-to-end response latency.
Implemented a Python/Flask REST API connecting voice input, LLM reasoning, retrieval, and voice output.
Independently built the LightRAG-based retrieval layer for company and product knowledge.
Supported Twilio call-flow testing and Linux/Nginx/SSL prototype setup.
The broader voice assistant pipeline was collaborative; the LightRAG retrieval layer was Matei’s independent contribution. The system was a prototype/foundation requiring further latency and reliability tuning before production use.
Prism is a Chrome extension for customizing websites users already visit. It uses AI-assisted restyling to change page appearance without asking a model to rewrite an entire webpage.
Uses a compact anchored DOM representation so smaller models can reason over live page structure.
Applies patch-based CSS and DOM edits instead of full-page rewrites.
Supports element picking for precise edits and visual style templates for quick direction.
Keeps saved versions/history with restore and delete functionality.
Clavis is Matei’s private AI operations layer for farcas.dev, controlled through Telegram and designed around tool use, scheduled tasks, site maintenance, and verification.
Connects messaging, tool use, scheduled tasks, maintenance work, and verification into one operational workflow.
Includes a separate public-facing assistant with approved website context only.
Uses rate limits, CORS restrictions, in-flight request protection, and prompt-injection boundaries.
Logs visitor questions and uploads them to Google Sheets for review, helping identify content gaps.
OpenAI Parameter Golf is a constrained language-model challenge with a 16MB artifact limit and a 10-minute training budget. Matei submitted an Encode-Think-Decode transformer variant designed for that constraint.
Used an Encode-Think-Decode architecture with U-Net-style skip connections.
Applied Int5 GPTQ quantization to reduce the model artifact size.
Reached 1.1169 BPB within the challenge constraints.
Matei also built a JavaScript visualizer to explain the BPB, training, architecture, and compression tradeoffs behind the submission so the result could be inspected more clearly than a raw PR alone.
IBM BOB Hackathon was a 48-hour AI-assisted build sprint where the original browser-customization prototype was created. That prototype later continued separately as Prism, while this card preserves the original competition context.
Built the original AI-assisted browser-customization prototype during the hackathon sprint.
The public submission and original repo remain linked as the source context for the event.
The certificate records completion, attendance, and successful submission of a solution using Opus, IBM, REST API, and OpenAI.
HBN² AML Challenge, 2024: a team machine-learning challenge for fraud and money-laundering detection using transaction datasets.
REST API design, API architecture/integration, backend development, backend services, third-party APIs, JWT/OAuth/API-key auth, and database design/management.
LLM integration, AI agents, prompt engineering, workflow automation, RAG, LightRAG, voice AI, AI-powered web applications, and open-source LLM/image model experimentation.
Linux administration, Debian systems, Raspberry Pi deployment, Proxmox, VM provisioning, Nginx reverse proxies, SSL setup, server maintenance, Linux server workflows, and network troubleshooting.
Python, JavaScript, SQL, Bash, HTML/CSS, Chrome Extensions, WebSockets, full-stack development, responsive web interfaces, automation scripts, scrapers, and technical documentation.
