Elfege Arthur Leylavergne

Over Height Vehicle Detection — Real-Time Safety System for NYSDOT

Single-handedly architected and delivered production-grade edge-computing safety system — real-time over-height vehicle detection via integrated LiDAR, VCA cameras, and automated traffic control. Owned full stack from application architecture to field hardware configuration.

• Architected fault-isolated multi-process system (5 data processors + Flask app) in single Docker container with 7-layer watchdog architecture; achieved 120+ days continuous uptime.
• Implemented VCA streaming optimization: 18 duplicate detections to 1 optimal output via confidence plateau detection, maintaining ~100MB memory footprint.
• Designed dual-sensor correlation: LiDAR (TCP, millimeter-accurate) + VCA (RTSP classification) with 5-second deduplication and UUID-based video matching.
• Built MQTT/WebSocket distribution for TMC dashboards; AVRO pipeline integration with Apache NiFi.
• Created Express.js admin server in isolated container with SSH proxy for secure TMC remote management.
• Developed automated field installer handling encryption, hardware integration (LiDAR, PTZ, SNMP relays), and system service configuration.
• Engineered PostgreSQL batch operations (100/batch) with 6-day automated retention under 24/7 write loads.

DMS Control API — Highway Dynamic Message Sign Command & Control

Sole architect and developer of a production command & control API for Dynamic Message Signs across multiple state DOT deployments. Reverse-engineered NTCIP 1203 protocol from MIB specifications and IRIS open-source reference to build a vendor-agnostic driver framework — no SDK access provided.

• Designed driver-based architecture with factory/registry pattern: manufacturer-specific SNMP drivers (Daktronics VFC-3000, generic NTCIP), 4-gate request router, 3-tier quirk resolution (device override > model template > defaults).
• Implemented full NTCIP 1203 message lifecycle: MULTI markup validation, CRC computation, message activation sequences, readback verification, and scheduled message management via APScheduler.
• Built interactive API dashboard with live endpoint testing, Mermaid-based architecture diagrams, and field-level contextual help across 15+ endpoints.
• Delivered 31 end-to-end tests, Kafka event streaming integration, and MIB-driven capabilities discovery — all from protocol specs with zero vendor support.

Device Status Monitoring System — Network Infrastructure Monitoring for UDOT

Architected production-grade network monitoring system for Utah DOT, tracking 2,400 devices across 750+ subnets on state fiber infrastructure.

• Achieved 90%+ cycle time reduction (54–60 min to 5–7 min) through device-type filtering and adaptive batching.
• Designed Python asyncio/aiohttp architecture with configurable worker pools (2,000 concurrent pings) preventing file descriptor exhaustion.
• Engineered APScheduler-based per-device-group scheduling with independent intervals (12/24/48 min cycles), 100% adherence.
• Integrated PostgREST streaming writes for real-time visibility; containerized via Docker Compose with AWS (ECR, Secrets Manager, SSO).
• Validated 120+ hours continuous operation, zero errors.

Real-Time Alert Routing System — Event Processing & Notification

Architected real-time alert processing system using Apache Flink and Kafka, dispatching personalized notifications via email, SMS, and WebSocket.

• Designed event-driven architecture processing complex trigger conditions across multiple event types.
• Integrated PostgreSQL with stored procedures for optimized operations; modular dispatch supporting SMTP and Twilio.
• Containerized with Docker Compose; implemented multiprocessing/multithreading for concurrency.

Video-Streaming Platform — Low-Latency DOT Camera Infrastructure

• Built core video-streaming platform: Flask + FFmpeg encoding/compression with WebRTC/HLS delivery.
• Implemented React monitoring interface for real-time stream health and metrics visualization.
• Automated deployment infrastructure via Bash scripting and CI/CD pipelines.

Front-End — DotStream React/Next.js Monorepo

Contributor to production Next.js 15 / React 19 / TypeScript monorepo (Turbo, pnpm workspaces) serving 15+ dashboard pages across multiple state DOT deployments.

• Integrated full OHVD real-time subsystem into the React platform: alert providers, WebSocket event streams, map panel hierarchies, and Kendo data grids.
• Contributed to migration from 5,000+ line vanilla JS monolith to modular React component architecture with Jest unit tests and Playwright E2E coverage.

Linux Infrastructure & Bash Automation — Personal Homelab

Self-built production-grade homelab on Dell PowerEdge R730xd (Proxmox VE), managed entirely via Bash automation. Bash is the primary tool for infrastructure operations — not a fallback, but the preferred layer of control.

• Administers Ubiquiti network stack (USW-Pro-48, UXG-Fiber gateway, 6× UAPs), SonicWall firewall, VLAN segmentation across 8 network zones.
• Built and maintains ~/.bash_utils — personal automation library: AWS SSO auth chains, SSH ProxyJump orchestration across 5 machines, PDF export pipeline, secret rotation, Docker lifecycle management.
• Runs 20+ self-hosted services (Proxmox VMs/LXC, NVR, Anamnesis, job tracker, Ollama inference) on isolated network segments.
• SSH ProxyJump chains: office → server → app1 → dellserver with per-host key management and .ssh/config templating.

Job Application Tracker — Full-Stack AI-Assisted Workflow System

Designed and built a full-stack application tracking system integrating AI generation, browser automation, and email parsing — a real engineering problem solved with production-grade architecture.

• FastAPI async backend, Next.js 15 / React / TypeScript frontend, PostgreSQL, Docker Compose; modular API with 15+ endpoints.
• Chrome Extension (Manifest V3) integrates with ATS platforms via Chrome Debugger Protocol (Input.dispatchKeyEvent) — required reverse-engineering react-select component internals that block all synthetic DOM events.
• Nightly APScheduler pipeline: SSH to remote machine, parses Thunderbird mbox for LinkedIn/CareerBuilder alert emails, extracts job postings, deduplicates against DB.
• Claude CLI integration: AI-generated cover letters, cached fit assessments, Anamnesis episodic memory for cross-session learning.
• Event dedup system: URL-hashed form interaction tracking with confirmation guards — prevents duplicate submissions.

Anamnesis — Vector-Based Episodic Memory for AI Instances

Designed and built a semantic memory system enabling AI instances to retain and retrieve contextual knowledge across sessions — grounded in epistemological theory of episodic memory.

• Architected FastAPI async backend with MongoDB Atlas Local native $vectorSearch (HNSW, 1024-dimensional embeddings) — zero cloud dependency.
• Built multi-model embedding pipeline (sentence-transformers) with automatic re-embedding on model change, checkpoint-based resumption, and JSONL bulk ingestion.
• Implemented cross-instance coordination: file crawler auto-ingests from handoff buffers, project histories, and CLAUDE.md files across networked machines.
• Integrated streaming chat with memory injection — semantic search before each turn, session persistence with rename history, three LLM backends (Ollama, Claude CLI, Claude API).

MOBIUS.NVR — Multi-Vendor Video Surveillance Platform

Unified NVR managing 17+ IP cameras across 4 vendors and growing (Python/Flask/Docker), with ONVIF/WebRTC/MediaMTX streaming at 200ms latency and Neolink bridging for non-ONVIF devices.

IoT & Home Automation — Custom Firmware, Robotics & Infrastructure

• Reverse-engineered iRobot Roomba 980 WiFi protocol (HTTPS/JSON API/Node.js) — first SmartThings/Hubitat integration, subsequently forked and extended by community developers.
• Thermostat Manager (Groovy): multi-device HVAC automation with narrow-AI learning features. Dashboard interface (JavaScript/HTML) for real-time IoT device control.
• KOBRA robot (ESP32): custom-built autonomous platform — dual motor+encoder closed-loop control, WiFi webserver, 3-DOF arm control, magnetic docking system; two full hardware generations.
• MEBO robot (ATmega2560 + ESP32/ESP8266): commercial toy reverse-engineered and extended across multiple iterations — custom motor driver, WebSocket interface, virtual joystick frontend, full firmware replacement.
• 6+ motorized window actuators deployed across apartment (A4988 stepper, H-bridge DC, and linear actuator variants) — each with custom C++ firmware + matching Groovy Hubitat driver.
• Door intercom automation: NodeMCU monitors analog voltage on LCD power line to detect doorbell ring; state machine drives TALK + DOOR relay sequence automatically — no hardware modification required.
• IR blaster network: split AC units (3 bedrooms + living room), Apple TV, Samsung TV — protocol-decoded and replayed via ESP8266 over MQTT.
• Environmental sensor network: motion (6 rooms), humidity, leak, water level, bed occupancy, gas sensors — all custom firmware on ST_Anything-adapted architecture with matching Groovy Hubitat drivers.