A multi-agent system that writes job-tailored resumes and cover letters that are grounded in your real experience, optimized for AI screeners, and quality-gated — every single claim traced back to a verified fact before anything exports.
Resume Talos takes a job description and a knowledge base of your professional history, then runs a fleet of specialized AI agents to analyze the role, score fit, retrieve the right evidence, draft and revise, simulate how automated screeners and recruiters will read the result, and hard-gate on groundedness before writing a single DOCX or PDF.
The organizing principle is anti-hallucination: a resume that invents a credential is worse than useless. So a Groundedness Verifier sits at the end of the pipeline and blocks export unless every factual claim traces to a specific knowledge-base fact.
 |
 |
Two layers of ATS simulation and a recruiter-triage verdict (left) · a grounded knowledge base every claim traces back to (right)
flowchart TD
JD(["Job description + knowledge base"]) --> A["JD Analyzer<br/>requirements + success signals"]
A --> F{"Fit Scorer<br/>cheap early filter"}
F -->|weak fit| STOP(["Stop — save the spend"])
F -->|good fit| K["Knockout + KB-gap detection<br/>disqualifiers surfaced honestly"]
K --> R["Retriever · pgvector<br/>pulls the right KB facts"]
R --> W["Resume + Cover-letter Writers<br/>grounded drafting"]
W --> SIM["ATS · recruiter · screener simulation<br/>how it will actually be read"]
SIM --> QC["Dual QC — Claude + Grok — + consolidator"]
QC --> V{"Groundedness Verifier<br/>every claim → a verified KB fact"}
V -->|unverified claim| W
V -->|all grounded| EX(["Export DOCX / PDF<br/>parseability + page-fit gated"])
classDef gate fill:#fde68a,stroke:#b45309,color:#111827;
classDef done fill:#bbf7d0,stroke:#15803d,color:#111827;
class V gate
class EX done
| Design decision | Why |
|---|---|
| Grounded generation with a hard verifier gate | Every claim in the resume/cover letter must trace to a KB fact ID. The Groundedness Verifier is a blocking gate — no export until it passes. This is what makes AI-written application material trustworthy. |
| Screener- and recruiter-simulation, not guesswork | Dedicated agents (ATS simulator, ATS-vendor-aware analysis, recruiter simulator, 7-dimension screener rubric) model how the document will actually be parsed and skimmed — then feed fixes back into revision. |
| Knockout detection up front | A KB-gap detector and knockout detector catch disqualifying requirements (missing clearance, cert, degree) before spending model budget drafting — and surface them honestly instead of papering over them. |
| Independent dual QC | QC Reviewer A (Claude Sonnet) and QC Reviewer B (xAI Grok) review independently; a consolidator merges them. Different model families = genuinely independent second opinions, not an echo. |
| Per-role multi-provider routing | Anthropic (Opus/Sonnet/Haiku), Google Gemini, and xAI Grok are each assigned to the roles they're best at, via a per-role model registry with cost tracking and prompt caching. |
| Cheap filters before expensive drafting | A Haiku fit-scorer runs first; market research is cached by company; a hard cap of 3 review iterations bounds spend. |
| Parseability-gated export | Output isn't just styled — a parseability checker verifies an ATS can actually extract the headings/sections, page-overflow trim keeps it to length, and a mandatory-content enforcer guarantees required items (e.g. specific certifications) survive every pass. |
| Agent | Role |
|---|---|
jd-analyzer |
Extract requirements, skills, and success signals from the JD |
fit-scorer |
Cheap early filter — is this role worth pursuing? |
knockout-detector · kb-gap-detector |
Flag disqualifiers and evidence gaps before drafting |
market-research |
Company culture/tone (Google-grounded), cached per company |
retriever · screening-similarity |
pgvector cosine retrieval of the right KB facts |
resume-writer · cover-letter-writer |
Draft and revise, grounded in retrieved facts |
screener · recruiter-simulator · ats-simulator · ats-vendor |
Simulate AI-screener + human-recruiter + ATS-vendor reading |
qc-reviewer · qc-consolidator |
Dual independent QC review + consolidation |
verifier · verifier-fix-suggester |
Groundedness gate: every claim → a KB fact, with fix routing |
cert-acronyms · questionnaire-helper · exemplars |
Cert normalization, application-question help, style exemplars |
Supporting subsystems: a KB pipeline (src/lib/kb/) for ingestion, chunking,
fact extraction, dedup, career-timeline and tenure reasoning; an export engine
(src/lib/export/) with three resume layouts (classic / executive / modern) in
both DOCX and PDF, plus the parseability, trim, and mandatory-content gates.
- Checkpoints: fit-score approval → long/short variant choice → market-research approval before cover-letter writing.
- Gates: both QC reviewers > 90 with no high-priority issues → early stop; max 3 iterations then escalate to an editable web view; groundedness verifier must pass before any DOCX/PDF is written.
- App: Next.js 16 + React 19 + Tailwind v4 + shadcn/ui
- Data: Drizzle ORM + Postgres + pgvector (1536-dim cosine, HNSW indexes)
- LLM: Vercel AI SDK v6 → Anthropic + Google + xAI (+ OpenAI), per-role model registry with pricing/cost tracking and a structured agent-run log
- Export: DOCX (
docx) + PDF (@react-pdf) in three layouts each
src/
├── app/ # Next.js App Router (dashboard, applications, knowledge-base, settings)
├── lib/
│ ├── agents/ # 20 LLM agent modules
│ ├── kb/ # ingest → chunk → extract → dedup → timeline reasoning
│ ├── export/ # DOCX/PDF layouts + parseability / trim / mandatory-content gates
│ ├── applications/ # application lifecycle, QC loop, versioning, export
│ └── models/ # per-role registry, cost tracking, embeddings
└── db/ # Drizzle schema + migrations
scripts/ # KB seed/maintenance + deterministic test & smoke scripts
Note on data: the knowledge base and contact details are user-provided and live in your database (not in this repo). Seed scripts under
scripts/show the shape; sample contact values are placeholders. Provide your own via the KB ingestion UI and settings.
pnpm dev # http://localhost:3200
pnpm exec tsc --noEmit # type check
pnpm db:studio # browse the DB