ShippedJun 2026 – Oct 2026Solo builder

Comfy Recruiter

An AI candidate-research agent and a hiring pipeline you can actually play — every signal cited, scored, and routed.

next.jslanggraph.jsclaudepostgres

The Field: the hiring pipeline as an isometric tactics board, candidates stacked in a pyramid with the strongest archetypes at the peak

Overview

A recruiting workspace and candidate-research agent for Comfy's frontend hiring pipeline. The agent resolves a candidate, fuses five external sources into one verified identity, scores the evidence against a ground-truth rubric with Claude, and routes the candidate onto the right hiring track — every signal backed by a citation and a confidence score.

The product wears an RPG skin: every candidate is a character sheet with stats, archetypes and status effects, and the whole pipeline is a tactics board — a hiring pipeline you can actually play. It has researched 770+ candidates and routed 268 recruits.

Candidate PII never leaves the local workspace: the data directory is git-ignored, writeback to the ATS is dry-run by default, and every research run is traced in a self-hosted Langfuse.

Media

The product tour: a siege cinematic opens into the live app — roster, character sheets, and the Field, with fictional demo candidates.

Comfy Recruiter landing page in pixel-art styling: 772 candidates, 771 researched, 268 recruits, with three sections — Recruit Roster, The Field, Admin Dashboard — 772 candidates researched and routed — styled as a game you can play.

A candidate character sheet with archetype match bars, stat lines, special abilities and status-effect flags — a fictional demo candidate — Every candidate is a character sheet: stats, class, abilities, risks.

Features

Recruit roster

Every candidate as an RPG character sheet: stat lines, archetype match (Canvas Mage, Graphics Warlock, Craft Paladin), special abilities, and status-effect risk flags.

The Field

The pipeline as a tactics board: 600+ candidates in review stacked into a pyramid by archetype strength, strongest at the peak, rejects forking off.

Cited, scored research

Five sources — Ashby, GitHub, People Data Labs, Exa, Tavily — fused into one verified identity with per-field provenance and a name-confidence level.

Safe by default

The agent composes ATS notes and tags but writes nothing unless explicitly enabled, and missing API keys degrade gracefully to stubs.

Technical

A six-node LangGraph.js pipeline inside a Next.js 16 app: resolve candidate → build identity → GitHub and web research in parallel → analyze → ATS writeback. The analyze stage is Claude with Zod-validated structured output — JD alignment, three archetype scores, flags, a recommendation, and a hiring path with priority tier.

Web research is filtered against identity anchor tokens to keep results on the right person, and scoring runs against a ground-truth rubric (JDs, pipeline paths, a stage-by-stage signal map) bundled verbatim into the prompt. Runs are traced in self-hosted Langfuse v3 on a local Docker stack.

Identity fusion with provenance

The ATS record, GitHub, and People Data Labs merge into one identity with per-field source attribution and a name-confidence level, so every downstream claim is auditable.

PII stays local

The candidate data workspace is git-ignored and lives only on local machines; the published demo uses fictional candidates throughout.