Job Search Agent: AI-Powered Career Pipeline

Before: Manual Career Page Browsing#

Filter Ordering as the Cost Strategy#

A naive pipeline would hand every one of the 58,807 weekly postings to an LLM for classification. At roughly $0.05 per job for careful extraction, that comes out to roughly $2,900 per week, or $150,000 per year, to classify postings that are mostly obvious rejects from the first sentence. The architectural pivot was to treat the LLM as the last filter, not the first. Free regex filters reject 94% of postings before any paid work begins. A SHA-256 deduplication pass removes another 16% of what remains. Only 312 postings reach Claude Haiku 4.5, and only 311 survive to the dashboard. The cheapest API call is the one you never make.

Three Claude Models, Three Roles#

Rather than one generalist model doing everything, the system assigns the right Claude to the right job. Haiku 4.5 handles the high-volume, repetitive enrichment work where the per-call cost has to stay near zero. Sonnet handles the rare, high-stakes code generation work where reasoning quality matters more than per-call cost. Claude Code handles the maintenance loop, where stateful context and tool orchestration matter more than raw model calls.

The Analyst (Claude Haiku 4.5) classifies seniority, extracts salary ranges, identifies requirements, formats descriptions, and decides whether each posting is a real software role. It also re-enriches on demand. Paste a job description into the dashboard’s edit dialog, hit Save, and Haiku re-analyzes it in about 2 seconds for $0.0003.

The Builder (Claude Sonnet) writes browser-based retrievers automatically through a multi-agent system built on the Anthropic Agent SDK. An orchestrator dispatches companies to three parallel worker agents. Each worker visits the career page, analyzes the DOM, writes a retriever, tests it in a subprocess, and self-corrects until the tests pass. The key was not clever prompting. It was a 1,466-line curriculum, packaged as a Claude Code Skill with reference docs, that teaches the agent how ATS platforms actually behave. 69 production retrievers, $49.68 total, zero manual edits.

The Maintainer (Claude Code) runs the self-healing layer. Custom hooks inject failure context into every new session before the first prompt. Custom slash commands orchestrate two-agent investigation-and-repair loops. The engineer types one command and three minutes later the system reports what broke, what it changed, and how many jobs are flowing again.

The Multi-Platform Reality#

Every ATS platform has exactly one quirk that will ruin the first attempt. Dayforce has persistent WebSocket connections that break page-load detection. Ashby has case-sensitive URL slugs where Jasper%20AI returns jobs but jasper returns nothing. Workday uses opaque tenant URLs where bb.wd3.myworkdayjobs.com maps to BlackBerry with no discoverable pattern. BambooHR has mojibake encoding that garbles diacritical marks. NeoGov uses /jobs/ instead of /job/, a two-hour debugging session for a single missing character.

The 11 platforms split across three tiers of difficulty:

• Tier 1 (Clean APIs): Greenhouse (177 companies), Lever (46), Workable (32), Pinpoint (4). REST or GraphQL endpoints return structured JSON.
• Tier 2 (Predictable SPAs): Ashby (1,558 companies) and Workday (15). Single-page applications with consistent DOM selectors, accessible via browser automation.
• Tier 3 (The Wild West): SAP SuccessFactors (4), SmartRecruiters (3), Dayforce (2), BambooHR (8), and one-off government portals. CGI forms, WebDynpro, legacy encodings, and pages that have not been redesigned since the early 2000s.

The 1,558 Ashby companies deserve a callout on their own. Ashby’s GraphQL API covers most of them, but a long tail of companies deploy Ashby as an embedded SPA without exposing the underlying API. Browser automation fills the gap, hitting the same data the user sees in their own browser.

Self-Healing as a First-Class Concern#

Companies change platforms, rebrand, and restructure their career pages. A static collection script on 1,975 companies will always have some subset failing at any given time. The architectural decision was to stop trying to prevent failures and instead respond to them automatically.

The retriever-debugger (Haiku, approximately $0.05 per investigation) diagnoses failures. It web-searches the company, fetches the current careers page, detects ATS fingerprints in the HTML, and produces one of four verdicts: ats_fix (slug changed within the same platform), platform_migration (company moved to a different ATS entirely), new_retriever (no API available, browser automation needed), or transient (temporary issue, skip and retry). There is no “deactivate” verdict, by design. Every company stays in the pipeline.

The retriever-fixer (Sonnet, approximately $0.30 to $1.50 per repair) takes the diagnosis and applies the fix inside a 3-iteration test loop. Database updates for slug changes. Full browser-based retrievers built from scratch when APIs disappear. A PostToolUse hook runs ruff check after every file write, catching lint errors in real time so the agent sees them on the next iteration.

The Monday morning workflow tells the story. Sunday night the pipeline runs. Monday morning the engineer opens Claude Code and the SessionStart hook injects “5 failures in 72 hours” into context before the first keystroke. Running /fix-all-retrievers kicks off the investigate-fix-test loop for each failure. Roughly 15 minutes later: 4 of 5 fixed, 1 logged for manual review. Running /pipeline-review confirms a 97 of 100 health score. Total spend: around $5.

A typical repair transcript, compressed to the interesting lines:

1
$ /fix-retriever jasper-ai
2

3
[debugger] Fetching https://www.jasper.ai/careers...
4
[debugger] Detected ATS fingerprint: Ashby (previous: Greenhouse)
5
[debugger] Verdict: platform_migration
6

7
[fixer] Updating companies table: platform Greenhouse -> Ashby
8
[fixer] Generating Ashby retriever for slug 'jasper'
9
[fixer] Test iteration 1: 0 jobs returned (slug case mismatch)
10
[fixer] Retrying with slug 'Jasper' (capitalized)
11
[fixer] Test iteration 2: 13 jobs returned, schema valid
12
[fixer] Committing retriever, updating health score
13

14
Repaired: jasper-ai (13 jobs flowing). Elapsed: 2m 47s. Cost: $0.41.

AI at Six Distinct Touchpoints#

AI appears across the system as a set of targeted interventions rather than a single monolithic feature. Each touchpoint has its own cost profile, latency budget, and failure mode.

1. Pipeline enrichment ($5.04 per week) — bulk classification of 312 candidate postings through Haiku 4.5.
2. On-demand re-enrichment ($0.0003 per job) — triggered when a user edits a description in the dashboard.
3. Job capture enrichment (background task) — structured metadata for manually-added postings.
4. Retriever builder ($0.72 per retriever, 69 built) — Sonnet inside the Agent SDK multi-agent system.
5. Retriever fixer (3-iteration cycles, $0.30 to $1.50 each) — Sonnet applying diagnosed repairs.
6. Retriever debugger (~$0.05 per investigation) — Haiku diagnosing failures before any repair is attempted.

The pattern that ties them together is deliberate model selection by cost profile. Haiku for the high-volume, lower-stakes work. Sonnet for the rare, high-stakes reasoning. No model is doing a job that a cheaper model could have done.

The Dashboard: Turning 311 Matches Into Something Usable#

311 qualified jobs per week are only useful if two applicants can actually work through them without stepping on each other. The dashboard turns the firehose into something manageable.

The core is an ag-Grid Enterprise v35 table with 35+ columns, but the whole point of Enterprise is not showing all 35 at once. The Columns panel lets each user toggle any column on or off, drag to reorder, group rows by company or seniority, and save the entire layout as a named preset. Each user gets their own persistent view.

Multiple applicants share the same job data but maintain separate views, favorites, notes, and application tracking. The application status pipeline tracks each job through 13 states:

1
New → Saved → Applied → Screening → Round 1 → Round 2 → Round 3
2
                              → Final Round → Offer → Rejected
3
                              → Withdrawn → Ghosted

Color-coded badges provide instant scannability across the grid.

On-demand AI enrichment. Editing a job description triggers an instant re-enrichment pass through Claude Haiku 4.5. Paste text from a company’s careers page, hit Save, and eight cells update simultaneously: seniority, salary range, requirements, employment type, and formatted description. Same model, same extraction, same $0.0003 cost as the batch pipeline, but triggered by a user clicking Save.

Jobs found outside the automated pipeline can be added through a manual entry form. After saving, Haiku 4.5 automatically enriches the posting with structured metadata.