unmpld

It's an interesting time to be job hunting. Beyond your network, finding jobs has meant browsing job boards. Those tend to hoard listing data behind increasingly sophisticated walls. Most of the companies who provide that data put it on their own websites, because the whole point is to attract applicants.

Displaced tech nerds with AI subscriptions are turning away from the job board walls and building their own job-hunting agents. This is one of them, built by Steve Drasco (physicist turned slopportunist) and Claude (Anthropic's AI). Steve steers, Claude builds. Once called vibe coding or AI slop, this kind of project may have matured to harness engineering. For some reason, everyone building these makes them multi-user. Nobody asked for that. It just happens.

The system scrapes company career pages, browsing them the way a human would, reading the HTML and using an LLM to understand what jobs are listed. Some companies expose structured APIs (Greenhouse, Lever, Ashby) which give cleaner data; the crawler detects these automatically and upgrades to them when available.

Found jobs are scored against the user's preferences using a mix of hard rules and LLM judgment, then presented on a review board. A conversational chat agent manages the entire workflow (adding companies to track, refining preferences, reviewing candidates) and silently learns the user's interests over time.

Deliberately, this is a hunter-gatherer. It finds and evaluates opportunities. It doesn't apply on your behalf. Auto-applying is a different problem with different trade-offs.

If the predictions about AI displacement are even half right, there will be a lot more tools like this soon. The job boards know it too.

Key numbers

How It Works

You rant. The agent listens, builds a profile, and acts on it.

System Architecture

Three pipelines, one brain, one database per human.

Crawl Pipeline

Every 4 hours, the crawler visits every tracked company's careers page. If the company uses Greenhouse, Lever, or Ashby, it pulls clean structured data. Otherwise it reads the raw HTML and asks an LLM what jobs are there. New companies start dumb (scraping) and silently get smarter (API upgrade) when the crawler figures out a better way in.

Structured sources

Post-crawl

New jobs get scored immediately. Anything above 70 triggers an email with the top 5.

Scoring

90 points of hard math. 10 points of LLM vibes. The code does salary, location, keywords, role fit. The LLM handles the squishier stuff: "is this actually interesting?" It never touches arithmetic.

Dimensions

Dealbreakers halve the score and auto-dismiss. You said "never defence contracts" and you meant it.

Chat Pipeline

You talk to it. It does things. 22 actions, three entry points, one planner that decides what you meant.

Entry points

Full chat pipeline

The planner sees everything: your jobs, preferences, profile, stats. If it decides to act, it acts then replies. If not, it just talks. Either way, it silently updates your profile.

Bot actions (22)

Implicit learning

Say "I'm into climate tech" and the bot quietly adjusts your scoring preferences without being asked. Every conversation makes the next crawl smarter.

Onboarding

First login, empty board. The bot notices and asks what you're looking for. By the end of the conversation, companies are tracked and preferences are set.

Agent Design

The rule: code for facts, LLM for judgment. If it's arithmetic, a lookup, or has clear rules, code does it. If it requires reading between the lines, the LLM does it.

Why not just LLM everything?

LLMs can't count, can't do reliable arithmetic, and hallucinate when bored. So salary conversion, score math, location lookups, and keyword matching are all plain JS. Stats get pre-computed and injected as facts. The LLM receives answers, not questions.

The learning loop

Chat → profile updates → scoring adapts → better results → chat again. No configuration screens. You just talk and the system quietly gets better at knowing what you want.

Safety

Bulk actions on more than 10 jobs need confirmation. The planner sees the full board before deciding anything. No YOLO mode.

LLM Layer

One abstraction, two tiers, multiple providers. Opus for anything that matters. Haiku for anything that doesn't. Every call is token-tracked so we know exactly where the money goes.

Database

SQLite, because nothing else was needed. One central DB for system stuff, one per user for their jobs and preferences. Created on first login, cached forever.

Central DB: unmpld.db

Per-user DB: data/user_{id}.db

Stack

Vanilla everything. No frameworks, no build step, no ORM. Two npm dependencies in production. Runs on a single cheap VM. The kind of stack where you can read the entire codebase in an afternoon.

Live at unmpld.com.