AI Agent & Infrastructure Engineering Intern (Graduates Only

Please note before applying: this internship is open to graduates only (degree conferred; current students are not eligible) and is fully on-site in Azusa, CA, including an in-person interview — candidates should be in the greater Los Angeles area or able to relocate.

This is not a prompt-writing role and it is not an ML-research role. You will build the AI layer our company runs on — agent harnesses, context pipelines, model routing, internal tooling — and the physical compute cluster it runs on. Software and hardware, both.

ELDÆON is building the world's first tactical UAP-detection network. Our sensor-fusion platform is engineered ground-up to detect what conventional defense systems were never designed to see — anomalous aerial phenomena that violate restricted airspace, including near nuclear facilities and military installations.

Our platform is:

• Modular — adaptable across rooftop, field, and networked regional deployments
• Multimodal — fusing optical, RF, environmental, biometric, and temporal data
• Real-time — live detection, correlation, and characterization of aerial phenomena

We are sensor hackers, reverse engineers, and deep-tech builders. We build what others won't.

We run two flagship sensing platforms —
DIONYSUS (multi-sensor UAP-detection enclosure) and NEMESIS (passive radar) — across dev units, provisioned field units, and live deployments. They generate a constant stream of bugs, stale data streams, config drift, and integration gaps. Today humans find those bugs and humans fix them.

That's the problem you're here to solve. We want the agents finding and fixing, and engineers reviewing.

We have the beginnings of an internal agent system. What it lacks is the harness, the context, and the reliability to be trusted unsupervised — right now a bad autonomous fix could take down a data pipeline, so we keep a human in the loop on everything. Getting past that trust threshold is the single highest-leverage problem at this company, and it's yours.

You will also build the compute that makes it possible. We're standing up our own inference infrastructure — distributed NVIDIA GB10 nodes and RTX PRO 6000 servers — so we can run open-weight models on our own data without sending it to a third party. You'll rack it, network it, and load models on it.

• Build agent harnesses and internal tooling — the scaffolding that lets agents work our codebase, sensor fleet, and docs reliably instead of one-off prompting
• Design context pipelines that give agents real awareness of ELDÆON: repos, commit history, sensor telemetry, dashboards, Confluence, Discord, and Jira
• Implement model routing across Amazon Bedrock and Open Router — pick the right model per task on cost, latency, and capability, with fallback behavior
• Work with agentic coding tools (Claude Code, Codex) as production instruments, not chat toys — loop engineering, evals, guardrails, and PR-gated autonomy
• Deploy and serve open-weight LLMs on our own hardware
• Build internal software for our business and engineering teams — voice-to-text capture, automated triage of sensor and dashboard anomalies, agent-authored PRs that a human approves before merge
• Define how we measure agent reliability, so we can justify expanding what runs unsupervised

• Build a distributed compute cluster from multiple NVIDIA GB10 systems — routing, network switching, cabling, and interconnect for pooled VRAM and distributed workloads
• Spec and assemble our own GPU servers around RTX PRO 6000 class cards
• Own the full stack: power, thermals, networking, drivers, orchestration, model loading, monitoring

Required

• Completed BS or MS in Computer Science, Computer Engineering, Electrical Engineering, or a related discipline (degree already conferred, current students are NOT eligible)
• Strong Python and Linux command line
• Demonstrated hands-on work with agentic coding tools (Claude Code / Claude Desktop / Codex) — beyond casual use; you've built something with them
• Real understanding of LLM API mechanics — tool/function calling, context windows, streaming, token cost, prompt caching
• Comfortable…