--- title: "Recruiting AI: production matching, evals, copilots" description: "AI matching for platforms that move millions of people into work. Vector retrieval, eval suites, multilingual career copilots, and the boring infrastructure behind it." source: "https://www.kensink.com/industries/recruiting-ai/" canonical: "https://www.kensink.com/industries/recruiting-ai/" --- ★ Industry vertical · Recruiting AI 06 service items 8-week engagements INDUSTRY · RECRUITING AI · CAREER & TALENT TECH # Career tech that survives the launch. AI matching for the platforms that move millions of people into work. Vector retrieval, eval suites, multilingual career copilots, and the boring infrastructure that keeps a recommender running at half a million seekers. [Start a project →](https://www.kensink.com/contact) [View case studies →](https://www.kensink.com/cases) Industry Career tech · Recruiting AI Scale 500k+ seekers benchmarked Coverage 64 districts · 15 verticals Stack Direct LLM · no orchestration \[WHAT WE HEAR ON THE FIRST CALL\] ## Three pains every career platform hits. We have watched these patterns in conglomerate-backed marketplaces, independent talent platforms, and university-affiliated career portals. The shape is the same. The fix is not a better demo. The fix is the boring infrastructure underneath the demo. Eval gates, hybrid retrieval, cost meters, and the discipline to refuse a model that does not pass. PAIN · 01 01 / 03 ### Match quality stalls at scale. What worked at ten thousand seekers stops working at five hundred thousand. Embeddings drift, BM25 alone misses intent, and one bad reranker tanks a market. ↓ How we fix it, below. PAIN · 02 02 / 03 ### AI reliability is a board conversation. Conglomerate-backed platforms get asked about hallucination, fairness, and uptime in the same meeting as revenue. Without an eval bar, the answer is always a story. ↓ How we fix it, below. PAIN · 03 03 / 03 ### The architecture is being decided this quarter. Early-stage platforms lock in vector stores, model providers, and retrieval shapes that they will live with for three years. The wrong call costs a rebuild, not a refactor. ↓ How we fix it, below. \[SIX SERVICE ITEMS · ONE TEAM\] ## Pick the recruiting-AI problem. We'll bring the build. Each item is a focused engagement: eight weeks, fixed scope, eval suite at handoff. Bundle two or three when the problem warrants. SERVICE · 01 / 06 Core product AI matching engine The retrieval + ranking stack your platform runs on. Built to survive the launch and the next million users. - Hybrid BM25 + vector retrieval on Postgres - LLM-scored final fit with explainability - Cold-start strategies for new seekers and new JDs PostgreSQL pgvector OpenAI Cohere SERVICE · 02 / 06 AI reliability Eval suite for matching The number that turns 'AI reliability' from a board worry into an engineering metric. Gated on every release. - Golden set of validated matches per market - Relevance, diversity, fairness scoring - A/B harness + drift detection on production traffic LangSmith Promptfoo TypeScript Python SERVICE · 03 / 06 Multilingual RAG Career guidance copilot A career chat that holds context across multi-turn conversations and switches between English and the local language without losing the thread. - RAG over career data, mentorship transcripts, campus content - Mixed-language embeddings tuned for the market - Prompt patterns that do not switch language mid-answer Anthropic pgvector TypeScript OpenTelemetry SERVICE · 04 / 06 Employer side Skill-based hiring tools Resume parsing, skill graphs, and explainable ranking that gives recruiters a 'why this candidate' for every match. - NER + skill taxonomy mapping from resumes and JDs - Bias-aware ranking with audit logs - Per-candidate 'why this fits' explanation surfaced to employers Python PostgreSQL OpenAI Zod SERVICE · 05 / 06 Operate at scale Production observability + cost Cost-per-match metering, latency p95 per market, and quality-drift dashboards. The CFO sees the same numbers as the on-call. - OpenTelemetry traces across retrieval + rerank + LLM call - Cost-per-intent budgets per tenant or market - Eval-as-monitor: regression alerts before the user notices OpenTelemetry Datadog Grafana Sentry SERVICE · 06 / 06 Early-stage wedge Architecture review One-week audit of your retrieval stack, eval gate, model selection, and deployment shape. Written decisions you can defend to a board. - Vector store: Postgres + pgvector or vendor SaaS - Model + provider mix with exit costs scored - Eval suite shape: what to build first, what to defer ADRs Postgres vLLM Cloudflare Most engagements bundle two: a build (01, 03, 04) paired with the discipline that keeps it shipping (02, 05). Bring the shape closest to your blocker. [Scope your engagement →](https://www.kensink.com/contact) Want to see the K-Framework discipline behind every item? [Read the K-Framework](https://www.kensink.com/k-framework). \[THE STACK · BY LAYER\] ## Boring infrastructure. Production results. We pick the simplest tool that survives the audit. Most of the time that means Postgres, the model API, and your existing infra. LAYER · DATA + RETRIEVAL ### Data + retrieval. The store, the index, the search PostgreSQL pgvector Redis ClickHouse BigQuery OpenSearch LAYER · MODEL LAYER ### Model layer. Embeddings, providers, fallbacks OpenAI Anthropic Cohere Embed Voyage Llama (self-hosted) vLLM LAYER · EVAL + OBSERVABILITY ### Eval + observability. The eval bar, the cost meter, the drift alarm LangSmith Promptfoo OpenTelemetry Datadog Grafana Sentry LAYER · BACKEND + TRANSPORT ### Backend + transport. Type-safe everything TypeScript Next.js Python FastAPI gRPC BullMQ tRPC Zod LAYER · MOBILE ### Mobile. iOS + Android, native or cross React Native Expo Swift Kotlin FCM APNs LAYER · CLOUD + DEPLOYMENT ### Cloud + deployment. Whatever your infra already runs Cloudflare Workers Cloudflare R2 AWS GCP Vercel Fly ✕ WHAT WE DO NOT SHIP ### Direct against the model API. The same way you integrate against Postgres. - ✕ No LangChain - ✕ No LlamaIndex - ✕ No agent framework - ✕ No orchestration vendor - ✕ No black-box ML platform \[ PROOF · WHAT THE STACK DELIVERS \] ## Numbers that hold up in the second meeting. MEASURED · WEIGHTED · 2024–2026 SCALE 500k + Seekers benchmarked on hybrid retrieval REACH 64 Districts of Bangladesh coverage BREADTH 15 Industry verticals in match graph LATENCY p95 / 200 ms Median match-API latency target RECRUITING AI · APPLIED K-FRAMEWORK ## Bring the matching problem. We'll bring the build. Eight weeks, fixed scope, eval suite at handoff. Direct LLM engineering on top of the K-Framework. Two Q3 slots remain. [Start a project →](https://www.kensink.com/contact) [Read the K-Framework](https://www.kensink.com/k-framework) CYCLE 8 weeks · problem to live OUTPUT Code · evals · runbook LICENSE Yours, MIT, no lock-in