Kargo Hiring Agent

A shortlist tool for Arjun, built strictly to the case: it reads a CV, scores the candidate against a rubric derived from Kargo's own 8 past hires (not just the job spec), and drafts an interview brief and a candidate email (invite or rejection). It never sends anything on its own. Every email waits for an explicit, per-candidate confirmation.

Scores are persisted to Supabase (candidates table) and a local dashboard.csv. CVs can be supplied from a local path or fetched directly from a GitHub repository.

Stack

Layer	Tech
AI scoring	Gemini 3.8 Flash (google-genai)
Database	Supabase — project tsgvfaiojilehdhpjrqj
Email dispatch	Resend (only on explicit send)
CV sources	Local file / folder, or GitHub repo path

Setup

pip install -r requirements.txt
cp .env.example .env   # fill in your keys
source .env            # or export each variable

Key variables:

Variable	Required	Purpose
GEMINI_API_KEY	for real scores	AI scoring via Gemini 3.8 Flash
SUPABASE_URL	no (default set)	Supabase project URL
SUPABASE_KEY	no (default set)	Supabase publishable key
RESEND_API_KEY	for send only	Email dispatch
KARGO_FROM_EMAIL	for send only	Verified Resend sender address
GITHUB_TOKEN	for private repos	GitHub PAT for score-from-github

Without GEMINI_API_KEY, the pipeline runs in mock mode (keyword heuristic) so you can verify the plumbing before wiring in a real key.

Commands

Score one CV

python3 src/pipeline.py score --cv path/to/candidate.pdf --role pm
# or --role spm

Writes:

• output/<id>_brief.md — interview brief
• output/<id>_email.md — draft email (invite or rejection)
• appends to dashboard.csv
• upserts a row to Supabase candidates table

Score a folder

python3 src/pipeline.py score-batch --dir path/to/applications/ --role pm

Fetch CVs from GitHub and score them

python3 src/pipeline.py score-from-github \
  --repo owner/repo \
  --path applications/pm \
  --role pm \
  --ref main        # branch, tag, or commit (default: main)

Downloads all .pdf, .docx, .txt files under that repo path, scores each, and writes the usual outputs. Needs GITHUB_TOKEN for private repos.

List scored candidates

python3 src/pipeline.py list
python3 src/pipeline.py list --role pm

Prints a summary from Supabase (score, recommendation, email sent status).

Send a drafted email

python3 src/pipeline.py send --candidate-id <id>

The ONLY command that calls Resend. Prompts for confirmation, sends exactly the draft already on disk, then flips email_sent=true in Supabase.

Files

File	What it does
src/cv_parser.py	Reads .docx, .pdf, .txt into plain text
src/redact.py	Strips name, email, phone, URLs before the AI sees anything
src/scorer.py	Gemini 3.8 Flash call — returns scored JSON against the rubric
src/db.py	Supabase writes: upsert_candidate, mark_email_sent, list_candidates
src/github_source.py	Downloads CV files from a GitHub repo path
src/emailer.py	Transmits a pre-written draft via Resend
src/pipeline.py	CLI: score, score-batch, score-from-github, send, list
prompts/scoring_rubric.md	Four-dimension rubric (ops 40%, ownership 25%, role-fit 20%, comms 15%)
data/job_descriptions.json	PM and SPM job descriptions

Supabase schema

candidates (
  id UUID PRIMARY KEY,
  candidate_id TEXT UNIQUE,   -- e.g. "cand-8f3a"
  role TEXT,                  -- 'pm' | 'spm'
  total_score NUMERIC,
  recommendation TEXT,        -- 'interview' | 'reject'
  dim1_ops_exposure NUMERIC,
  dim2_ownership NUMERIC,
  dim3_role_fit NUMERIC,
  dim4_communication NUMERIC,
  dim1_evidence TEXT, dim2_evidence TEXT,
  dim3_evidence TEXT, dim4_evidence TEXT,
  interview_brief TEXT,
  email_subject TEXT,
  email_body TEXT,
  source_file TEXT,
  email_sent BOOLEAN DEFAULT FALSE,
  scored_at TIMESTAMPTZ DEFAULT NOW()
)

What's not built (on purpose)

• No autosend — every email is drafted, never dispatched, until send is run by hand.
• No scoring on name, photo, college, gender, or age — redact.py removes these before the model sees anything.