FoundingAppliedResearchEngineer

WHY THIS ROLE EXISTS Foundation models are commoditizing. Defensibility comes from specialized models, proprietary training signals, and evaluation ownership. Every applied AI company we benchmark against like Decagon, Harvey, Sierra, Cursor has already moved. The window to claim frontier applied AI for revenue is closing in the next few months. Rox is in market. We run agents against enterprise data at scale, every day. We see exactly where research meets production and where the data is dirty, state is changing, and being wrong costs (a lot of) money. The Applied Research team exists to close that gap permanently. WHAT THIS TEAM WORKS ON Four problems we care about right now: Cost-efficient inference for Clever Columns. Distill a Rox-trained model from frontier teachers so per-account enrichment runs at 1/20th the cost without quality loss. Ships first. Doesn't require trajectory attribution. Signal classification across the public knowledge graph. A small, fast classifier that distinguishes genuine buying signals from noise across the news, jobs, and filings corpus we already ingest at scale. Powers Recommended Next Moves and Auto Prospecting. Cleanest data subset. Personalization grounding and hallucination detection. A reward model that catches fabricated prospect context in Sequences in real time. This is the most underrated production failure mode in outbound AI. Trained on cross-customer consensus edits. Sequencing policy under sparse, delayed rewards. Offline-to-online RL on multi-touch trajectories with intermediate signals as proxies for terminal outcomes. Long-horizon flagship. Hard. [Depends on trajectory instrumentation in progress with Platform Eng.] These are not benchmark problems. They have real SLAs and real customers depending on them. WHAT YOU'LL DO - Design and run research programs tied directly to the four above. - Build evaluation frameworks that measure trajectory quality, not just final output, because most eval infrastructure measures end