Full-Stack Developer

Company Description ARCA, through its ARCATRANS division, is a logistics and transportation company with headquarters in Toronto, Ontario, Canada, and Mountainside, New Jersey, United States. The company primarily serves customers across Ontario and the East Coast of the U.S., providing reliable and efficient freight solutions. ARCATRANS offers a full range of services including LTL (Less Than Truckload), FTL (Full Truckload), FCL (Full Container Load), LCL (Less Container Load), local freight, and air freight. Team members at ARCA contribute to technology and operational improvements that support the movement of goods across borders and regions, helping customers optimize their supply chains. Role Description This is a full-time, on-site role based in Yerevan, Armenia, for a Full-Stack Developer focused on building and improving ARCA’s AI automation product for logistics operations. The role involves designing, developing, and maintaining web applications that integrate AI-driven features to optimize routing, load planning, and freight management. New product being built for the North American trucking industry, starting with one real customer already live: a cross-border freight carrier operating between Toronto and the US Northeast. Connects to that Gmail with a single OAuth click, reads every message, classifies it, routes it to a named owner in under a minute, and puts the carrier's own lane history and rate floors on the screen while they reply. You would be the first engineer on it. There is a complete written build plan, an interactive UI mockup that serves as the specification for the screens, and a founder who has done the data work and will be available daily — but who is not the builder. You would be the builder. Build Gmail ingest across six mailboxes (OAuth, watch/push plus a periodic reconcile sweep), with a coverage receipt on every run: messages in versus messages processed, with loud failure on any gap. A silently missed message must be impossible.Build the classification and extraction pipeline: offer / quote request / tender / status request / document request / customs / driver SMS / vendor notice / mass blast. Extract lane, dates, equipment, pallet count, weight and reference numbers, each with a confidence score and provenance, and a human-review queue below threshold.Deduplicate broker mass-blasts and stitch threads across mailboxes — the same request arriving five times into three inboxes must become one item.Implement territory routing that assigns a named human owner in under 60 seconds.Implement the thread-state ledger: two single-keystroke gestures (Mine / Done) that never delete anything, plus automatic reopen when a counterparty replies.Implement the ownership ledger and a 30-minute silence clock with escalation, as server-side scheduled jobs.Implement the automatic internal handoff chain, reusing the team's existing email subject conventions so they do not have to learn anything new.Build a settlement-PDF parser over weekly files going back to January 2020, and populate the money data model.Backfill lane memory from the historical order book.Build the owner's Command Deck: warning feed, pending decisions with age clocks, rules with fire counts. Qualifications You have shipped a production system that ingests messy external data — email, documents, EDI, scraping, or similar. Not only clean-API CRUD applications. This is the single most important requirement.Strong backend engineering in a typed stack: Node.js/TypeScript or Python/FastAPI, PostgreSQL, and a real job queue. You have built scheduled and recurring jobs that had to be idempotent and had to survive restarts.Practical experience with LLM extraction pipelines, and appropriate skepticism about them. You work with structured outputs, you write evaluations, you set confidence thresholds, and you design the human-review queue before you need it. "The model will handle it" is not an answer we accept.Reliability engineering instinct: reconcile loops, audit logs, degraded-mode behaviour. You raise failure modes before you are asked about them.Front-end capability sufficient to build a keyboard-first single-page application from an existing mockup. You do not need to be a design specialist — the mockup is the specification.Working English, written. All communication, documentation and code review is in English.Substantially full-time availability.