Elfege Arthur Leylavergne, Ph.D.

Trucks exceeding bridge clearance damage infrastructure and endanger traffic below. The system must detect within milliseconds, correlate multi-modal sensors, and never falsely trigger — a missed event is as unacceptable as a false one.

Sole architect. Fault-isolated multi-process architecture with layered watchdogs. Dual-sensor correlation — LiDAR plus video analytics — with intelligent deduplication. Detection pipeline optimized for tight latency budgets on modest edge hardware.

• 120+ days continuous uptime in production.
• Six-subsystem integration delivered by a single engineer.
• Production-deployed at a State DOT partner site.

A State DOT client needed real-time health of thousands of field devices distributed across hundreds of subnets. The previous polling architecture could not keep up; stale data was the operational norm.

Architected async monitoring end-to-end. Asyncio pools with adaptive batching, APScheduler per-group with adaptive intervals, PostgREST streaming writes to a staging table, trigger-based synchronization between dynamic and historical stores. The 4 ms end-to-end latency target was set at design — I identified the bottlenecks before writing the first async primitive. Deployment scripts refuse to boot if the OS ulimit is insufficient, and guide the operator to the correct configuration before anything starts.

• 90%+ cycle-time reduction over the prior architecture.
• Extended zero-error operation across thousands of devices.
• 4 ms API latency — deliberate, not emergent.

Highway Dynamic Message Signs speak NTCIP 1203 over SNMP, but every vendor interprets the MIB differently and proprietary SDKs are gated behind vendor contact. Traffic operations cannot wait for sales cycles.

Sole architect. Reverse-engineered NTCIP 1203 directly from MIB specifications to build a vendor-agnostic driver framework covering Daktronics VFC-3000, PCMS-548, and others. Full protocol lifecycle: MULTI validation, CRC computation, activation sequences, readback verification. Driver interface designed so adding a new vendor is a schema, not a rewrite.

• 15+ REST endpoints for full DMS command & control.
• Vendor-agnostic driver architecture — new signs plug in as data, not code.
• Delivered under a fast turnaround (~2 months) with no vendor SDK access.

Large language model instances forget between sessions. Conversations that should accumulate — architectural decisions, user preferences, hard-won context — dissolve. The standard workaround (stuff everything into the prompt) does not scale past a few turns.

Built a persistent episodic memory layer. 1024-dimensional sentence-transformer embeddings indexed in MongoDB, with semantic search, priority weighting, and time-decay boosting. Exposes a REST API over LAN: any Claude/LLM instance can write significant episodes and retrieve relevant ones on startup. In daily production use across a fleet of 20+ agent instances.

• 1024-dim sentence-transformer embeddings, MongoDB vector search.
• Multi-instance persistence: agents share continuity across machines & sessions.
• Daily use in real workflows — not a demo.

Traffic operations span many roles and agencies, each subscribing to different event classes with different urgency and routing. A naive fan-out wastes bandwidth and bores dispatchers into ignoring the feed; a single firehose cannot serve a multi-tenant platform.

Designed a multi-user subscription alert model on a Kafka / Flink backbone. Kafka ingests raw subsystem events (device status, detections, alarms); Flink carries the stateful stream processing and rule evaluation per subscriber. Alerts dispatch over email, SMS, and WebSocket to the exact audience their topology declares.

• Real-time dispatch across email · SMS · WebSocket channels.
• Multi-tenant subscription topology — per-user, per-role, per-event-class routing.
• Stateful stream processing on Flink over a Kafka event backbone.

The operator surface for dotstream's traffic-management platform: dashboards, a GIS map with real-time event feeds, and live video. The legacy JavaScript monolith had crossed 5,000 lines and could no longer absorb new subsystems cleanly.

Team contributor on the React refactor from the legacy monolith into a typed Next.js / React monorepo, and participant in the early CI/CD implementation. Built the operator-facing components for the safety subsystem I had architected on the backend — dashboard, map viewer with real-time events, and live video integrated against the dVS video-management backend I also wrote.

• Vertical ownership: same engineer on backend and UI for OHVD and dVS.
• 5,000+ line JS monolith lifted into a typed Next.js / React monorepo.
• Contributor to the platform's early CI/CD implementation.