Yusi Zhang

Summary

Tech Lead Manager at LinkedIn shipping 0-to-1 agentic AI products at startup pace. Currently leading two concurrent strategic bets — Interview AI (team of 8) and Expertise Graph (team of 10) — aimed at verifying candidate technical expertise at scale and growing LinkedIn's hiring and staffing-agency businesses. Over the past year my teams ramped three agentic AI launches to 100% production. Eleven years building large-scale data and AI platforms; previously Staff Engineer owning the pipeline that ingests 15M+ jobs/day.

Languages & systems: Python · Java · Kafka · WebRTC / WebSocket · Azure Communication Services · Realtime voice AI (GPT-realtime, STT/TTS) · LLM agents · prompt engineering & fine-tuning · LLM evaluation · distributed / multi-datacenter systems

Experience

• Shipped three 0-to-1 agentic AI products to 100% production in the past year:
  • Agentic Candidate Sourcing — agent that interprets hirer intent in natural language, semantically retrieves candidates, and applies an in-house fine-tuned LLM to score candidate × job fit on behalf of the hirer.
  • Agentic Outreach (InMail) — agent that drafts personalized outreach messages per candidate, conditioned on profile and job context.
  • Candidate-evaluation model upgrade — lifted in-house fine-tuned LLM accuracy from 60% → 90% on a 50K+ human-annotated golden dataset, driving customer CSAT up +0.5 on a 5-point scale.
• Now leading two concurrent 0-to-1 AI bets, both designated strategic for LinkedIn's next chapter; ~18 engineers across two orgs.
• Interview AI (team of 8) — architect. Rebuilding the voice-AI interviewing stack from a client-side prompt model to a server-side agent loop on Azure Communication Services + Realtime GPT. Selected and integrated WebRTC, WebSocket bridging, function-calling LLMs, and Azure Blob → durable storage under hard security and Python-at-scale constraints. Designed a STT/TTS-hybrid cost-reduction path that cuts per-interview inference cost by an order of magnitude.
• Expertise Graph (team of 10) — strategist. Translated an open-ended business goal — grow LinkedIn's staffing-agency margin — into a phased technical roadmap (member consent, evidence collection via interview transcripts, expertise synthesis, capability store, discovery) to verify candidate technical skill at scale.
• Cross-product loop: wired Interview AI evidence into the Expertise Graph pipeline so every AI interview produces a synthesized expertise signal that flows back into sourcing and evaluation — the strategic thesis behind running both projects together.
• Hiring across both teams; instituted on-call retros, weekly written status, and a tech-talk rotation to keep operational discipline at the pace 0-to-1 prototyping demands.

• 6× faster ingestion (~250 ms → ~40 ms p95) and ~90% fewer false errors — designed and shipped the post-ingestion processing framework that became the standard pipeline for downstream consumers of LinkedIn's job data.
• 15M+ jobs/day ingested with zero-touch site failover — led multi-datacenter ingestion design and rollout, doubling system capacity and removing manual intervention during disaster recovery.
• ~99% precision on closed-job detection at 4M jobs/week, taking down ~114K stale jobs/week and cutting member apply-clicks-per-reported-job from 79 to 67. Tech lead for a 4-engineer US/India team.
• Designed the agent platform behind LinkedIn's SMB hiring product — event-driven multi-agent workflow + quality/cost/throughput thresholding framework + evaluation infrastructure (golden datasets, precision/recall tracking, ramp criteria) used to certify each model release.

• ~95% lower Kafka consumer lag and ~70% faster processing via job-based partitioning and a new mining-task lifecycle — eliminated a long-standing hot-partition issue and took all job-source refresh from 72+h to 24h.

• Core contributor to the Job Ingestion pipeline; authored designs and reusable components later adopted by adjacent teams.

• Built and operated ingestion, provisioning, and CI services in Oracle's Big Data Cloud platform.

Education

Carnegie Mellon University — M.S., Information Technology

Zhongnan University of Economics and Law — B.Mgmt., Information Management & Information Systems

Languages: English, Mandarin (both native/bilingual).