AI Agent for Candidate Discovery for Large Scale Events

An AI research agent that turns a plain-language brief into a ranked, verified shortlist of real people built for a client in the experiential events space.

What it does: A user describes what they're looking for in a chat interface. The system extracts that into a structured, editable criteria spec, then kicks off a durable, multi-provider deep-research pipeline that plans search queries, sweeps the web in parallel, extracts and deduplicates candidates, enriches each profile with verified evidence, and scores/ranks the results typically surfacing 50-100 qualified candidates per run. Results stream live into an interactive results panel as they're found, and export to styled Excel or CSV.

Highlights:

Multi-provider AI orchestration coordinates OpenAI, Anthropic, xAI (Grok), and Exa web search in a single pipeline, with each model assigned the role it's best at (query planning, extraction, verification, scoring). Model IDs, roles, and pricing are centralized in typed config registries.

Durable, crash-safe workflows long-running research runs (20-30+ minutes) are built on Vercel's Workflow DevKit. A server restart or redeploy mid-run resumes cleanly: completed steps replay from persisted state, and all writes are idempotent so nothing double-bills or duplicates.

Real-time streaming UX candidates appear in the results panel the moment they're discovered, with enrichment filling in rows live over Supabase Realtime; a chat assistant can answer follow-up questions grounded in the actual run results.

Cost governance built in every model call is usage-tracked and priced; runs have hard cost caps and pause gracefully rather than overspend. No API spend happens until the user explicitly confirms the research criteria.

Trust and safety in the results verification gates with cited evidence, flagged-not-dropped handling for unverifiable candidates, and a strict no-fabricated-links rule: URLs only enter the dataset from real search/fetch results.

Quality engineering Zod schemas as the single source of truth for every data shape crossing a boundary, versioned prompt modules hashed per run for reproducibility, a calibration/benchmark harness for scoring quality, and a Vitest suite with all provider calls mocked.

Stack: Next.js 16 (App Router, Server Components) · TypeScript · Vercel AI SDK v7 · Vercel Workflow DevKit · Supabase (Postgres, Realtime, Auth with RLS) · Tailwind CSS v4 + shadcn/ui · Zod v4 · Vitest