Kairos

"Apply less. Land more." · Team Exitcode0 · Cursor Colombo Buildathon 2026

Kairos is a WhatsApp-first AI career agent. The conversational agent Kairo (OpenClaw + Baileys) coordinates job alerts, tailored CV generation, and cover letters. The web dashboard (Next.js + Supabase realtime) handles onboarding and tracking. Job matching and content generation use the MiniMax API; data lives in Supabase with pgvector for semantic matching when enabled.

---

Monorepo layout

Path	Purpose	Typical deploy
apps/web	Next.js dashboard (Tailwind + shadcn/ui)	Vercel
apps/poller	15-minute job polling worker	Railway
apps/cv-pipeline	CV YAML → PDF (RenderCV) + MiniMax content	Railway
packages/db	Supabase migrations, types, seeds	—
packages/minimax	Shared MiniMax client (scorer, CV, posts)	—
packages/matcher	Keyword filter, scoring orchestration, pgvector hooks	—
packages/config	Shared env validation, constants, types	—
agents/kairo	OpenClaw agent + Baileys WhatsApp runtime	Railway

---

Product brief

The problem

The modern job market rewards speed over qualification. Many candidates discover listings hours after posting and spend hours tailoring documents—by then, strong roles are already screened.

The solution

Kairos monitors job sources on a schedule, scores listings against a candidate profile (MiniMax + optional pgvector pre-filter), and sends an instant WhatsApp alert on strong matches. The user can reply with a single digit; Kairo generates a tailored CV (PDF) and optional cover letter in-thread. The dashboard updates in realtime via Supabase.

Demo script (wow moment)

1. WhatsApp alert: e.g. "87% match — Junior Full-Stack Developer @ Wise. Posted 6 mins ago."
2. User replies: 1
3. ~60s: Kavindu_Perera_Wise_2026.pdf in the thread.
4. Kairo asks about a cover letter; user says Yes → text reply.
5. Web dashboard shows the application logged without a manual refresh.

Differentiators

Typical tools	Kairos
Email digests (slow)	WhatsApp push (minutes after post)
Generic templates	AI-tailored CV per job
Browser/app context switching	Primary flow in WhatsApp
No match score	0–100 score + fit reasons
Manual tracking	Auto-logged dashboard

---

Technical architecture

Layers

• User: WhatsApp (Baileys on a dedicated device) + Next.js dashboard (onboarding, jobs, tracker, drafts, sources).
• Backend:
  • Kairo (OpenClaw): parses WhatsApp intents, triggers CV/cover flows, orchestrates MiniMax where needed.
  • Poller: cron (~15 min) — LinkedIn guest API, Greenhouse, Lever, Remotive; dedupe; matching; alerts for score ≥ threshold.
  • CV pipeline: (profile, job) → MiniMax JSON → RenderCV (EngineeringResumes theme) → PDF.
• Data: Supabase + pgvector; embeddings optional for similarity pre-check.

Matching engine (priority)

1. Keyword pre-filter — skills / target_roles vs title & description (cheap).
2. MiniMax scoring — structured { score, reasons[] }; threshold default 70.
3. Batch scoring — multiple jobs per call to reduce round-trips.
4. pgvector — cosine similarity pre-check before full scoring (when time permits).
5. Cache — avoid re-scoring same (user_id, job_id).

WhatsApp flow (Kairo)

Baileys message → OpenClaw intent
  • "1" / "yes" / "generate cv" → CV pipeline
  • "2" / "details" → job from Supabase
  • "3" / "skip" → skip job for user
  • cover-letter context + "yes" → MiniMax cover letter
  • else → Hermes / fallback conversational reply
→ Baileys send → write state to Supabase → dashboard realtime

Baileys: install only @whiskeysockets/baileys (verify package name; typosquat risk). Session dir: WA_SESSION_PATH (see .env.example). Implement reconnect for idle drops.

CV pipeline

1. MiniMax: CV content JSON (summary, experience[], skills[], projects[]).
2. Merge into RenderCV YAML (theme: EngineeringResumes).
3. PDF binary → Baileys document; log cv_version_url on applications.

Fallback: if RenderCV fails or times out (~45s), serve a pre-rendered demo PDF.

Supabase schema (reference)

• users — phone unique, timestamps.
• profiles — user skills, roles, JSON projects, optional embedding vector(1536).
• jobs — source, title, company, location, url, description, posted/fetched times.
• applications — user/job FKs, match_score, match_reasons[], cv_version_url, cover_letter, status.
• sources — integrations (Notion, Slack, etc.) with config JSONB.

Enable: CREATE EXTENSION IF NOT EXISTS vector;

Environment variables

See .env.example. Minimum: MiniMax keys, Supabase URL + keys, WA_SESSION_PATH, polling URL/interval, match threshold, NEXT_PUBLIC_APP_URL, RenderCV theme.

Deployment (recommended for hackathon)

• Vercel: apps/web
• Railway: agents/kairo, apps/poller, apps/cv-pipeline
• Supabase Cloud: database + realtime
• Tunnel (ngrok / Cloudflare): fallback if hosted workers are blocked

---

Dashboard screens

1. Onboarding / profile — uploads, parsing status, editable profile, WhatsApp number, activate CTA.
2. Job matches + tracker — live feed with score badges; application table; Supabase realtime; optional 7-day follow-up nudge.
3. LinkedIn post drafts — three variants (professional, conversational, technical); copy only, no auto-post.
4. Sources & integrations — cards with status (Notion + WhatsApp live where applicable; others "coming soon" but polished).

---

Team ownership (sprint reference)

Member	Focus
A	Next.js UI, Supabase realtime
B	Poller, LinkedIn guest API, schema + APIs
C	Kairo: OpenClaw, MiniMax wiring, Baileys flows
D	Deploy, RenderCV, parsers, demo data, rehearsal

---

Demo persona seed: Kavindu Perera

Pre-seed in Supabase before the clock. Example JSON (abbreviated):

{
  "name": "Kavindu Perera",
  "current_role": "Junior Full-Stack Developer",
  "experience_years": 1.5,
  "employer": "PayEase LK",
  "skills": ["React", "Next.js", "TypeScript", "Node.js", "PostgreSQL", "Supabase", "Tailwind CSS"],
  "target_roles": ["Junior Full-Stack Developer", "Frontend Developer", "Full-Stack Engineer"],
  "work_preference": "Remote",
  "career_goal": "Land a remote role at a product-focused international startup paying in USD.",
  "demo_target_job": "Junior Full-Stack Developer @ Wise",
  "expected_match_score": 87
}

The PayEase → Wise fintech angle is intentional for a believable high match in demos.

---

Risk register (summary)

Risk	Mitigation
LinkedIn rate limits	Cache last response; fall back to Greenhouse/Remotive
Baileys session drop	Reconnect logic; dedicated device for demo
RenderCV slow/fail	Static PDF fallback
MiniMax latency	User-facing "tailoring…" message; timeout + retry
Malicious npm typos	Only @whiskeysockets/baileys
Bad JSON from LLM	Validate + retry

---

Open decisions

1. Hosting: Railway vs tunnel mix for workers.
2. MiniMax embedding model for 1536-dim vectors (if using pgvector on profiles).
3. Batch size for multi-job scoring (start ~5).
4. Official judging criteria / pitch script.

---

Kairos — Apply less. Land more.
Agent: Kairo (OpenClaw + Baileys).
Built by Exitcode0 · Cursor Colombo Buildathon 2026