AutoJob-Agent

AutoJob-Agent is a modular, config-driven job-search automation pipeline:

• V1 Producer: fetches job postings from multiple ATS sources and maintains a single source of truth.
• V2 Feeder/Scoring: applies hard gates + semantic matching to triage jobs and outputs Apply/Maybe/Skip lists.
• V3 Local Agent (Ollama): generates human-readable Markdown job cards locally (not run in GitHub Actions).

This repo is designed so that V1 + V2 can run reliably in GitHub Actions, while V3 is intentionally local-only.

---

Repository structure

scripts/                 # V1 Producer scripts (fetch + status updates)
config/targets.json       # V1 targets and filters
data/                     # V1 outputs: jobs.json, state.json, today/backlog, archive/

v2/                       # V2 scoring + triage
  config/                 # scoring.yaml, profile.md (input for V3 too)
  scripts/                # score_jobs_v2.py, triage_v2.py
  data/                   # scored_jobs.json, emb_cache.json (optional)
  output/                 # candidates.json, apply.md, maybe.md, skip.md

v3/                       # V3 local-only job cards (Ollama)
  agent/                  # run_ollama_agent.py + prompt/template
  cards-samples/          # curated sample cards committed for demo
  cards/                  # generated cards (typically local output; often gitignored)

---

V1 Producer (fetch)

Inputs

• config/targets.json (companies + ATS settings + filters)

Outputs (single source of truth)

• data/jobs.json
• data/state.json
• additional reports:
  • data/jobs.md
  • data/jobs_today.json / .md
  • data/jobs_backlog.json / .md
  • data/archive/

Run locally

python scripts/fetch_jobs.py

Status behavior (important)

data/state.json stores job_uid -> status.

Jobs with status in applied / ignored / closed are hidden globally at the V1 output layer, so they will not reappear in V2 and won’t be repeatedly pushed.

---

V2 Feeder/Scoring (triage)

Inputs

• data/jobs.json

Outputs

• v2/data/scored_jobs.json
• v2/output/candidates.json (Top-N candidates for downstream use)
• v2/output/apply.md
• v2/output/maybe.md
• v2/output/skip.md (skip must include reasons)

Run locally

python v2/scripts/score_jobs_v2.py --config v2/config/scoring.yaml
python v2/scripts/triage_v2.py     --config v2/config/scoring.yaml

Scoring approach (high level)

• Hard gate rules (must-pass constraints)
• Semantic match using sentence-transformers all-MiniLM-L6-v2 cosine similarity
• Optional embedding cache for repeat runs

---

V3 Local Agent (Ollama job cards)

V3 requires local Ollama installed.
V3 runs locally only; not in GitHub Actions.

V3 reads the top candidates and generates a Markdown “Job Card” for each candidate, including match points, risks, skill gaps, resume-tailoring suggestions, and interview prep notes. The output is constrained to evidence from:

• v2/output/candidates.json
• v2/config/profile.md

Install Ollama (system-level, one time)

Install Ollama on your machine, then pull a small model that fits your available RAM. Example:

ollama pull qwen2.5:3b

Generate cards locally

python v3/agent/run_ollama_agent.py --model qwen2.5:3b

Limit generation (recommended for speed):

python v3/agent/run_ollama_agent.py --model qwen2.5:3b --limit 20

Inputs

• v2/output/candidates.json
• v2/config/profile.md

Outputs

• v3/cards/<job_uid>.md (generated cards)
• v3/cards-samples/ contains a few curated sample cards committed for demonstration.

---

GitHub Actions

GitHub Actions is configured to run V1 + V2 on the default branch and push updated outputs back to the repo.

V3 is excluded by design because it depends on local Ollama and is intended for local use only.

Local development setup

Recommended:

• Python 3.11 + virtual environment (.venv)

Typical workflow:

1. Run V1 fetch (or let Actions run it)
2. Run V2 scoring + triage
3. Optionally run V3 locally to generate cards for top candidates

Notes on tracked vs local-only outputs

• v3/agent/.cache/ stores local incremental indexes/logs and should not be committed.
• v3/cards/ is usually treated as local output (can be committed once for demo, but not recommended for frequent updates).
• v3/cards-samples/ is intended for curated examples.