back to work
Case studySolo build · 2026

Dossier

Quality-first agentic job-search SaaS

~$0.04
per full pipeline run
79
companies hand-scored
~60%
jobs dropped pre-LLM at $0
3-pass
self-evaluating resume gen

Pipeline

Profile + personaRule filter (−60% at $0)Scored discovery · 79 cosCompany + market intel3-pass LaTeX resumeSSE → live UI

The problem

Job search is high-effort and low-signal: hundreds of listings, most irrelevant, and tailoring a resume per role eats hours. I wanted a system that finds, scores, and researches roles autonomously — then produces an ATS-ready resume — for cents per run, not dollars.

Approach

01

8-agent autonomous pipeline

Persona Builder, Job Discovery, Watchlist, Company Intel, Gap Analysis, Market Intel, Resume Agent and Referral Finder — each a bounded agent with typed inputs and outputs, composed into one run.

02

Cost-first model routing

A pre-LLM rule filter drops ~60% of jobs at zero cost before any model call. Cheap models triage; expensive models only finish the shortlist — keeping a full run near $0.04.

03

3-pass self-evaluating resumes

Claude Sonnet tailors → Haiku critiques → Sonnet revises, emitting ATS-optimised LaTeX. A self-evaluation loop replaces one-shot generation.

04

CLI → multi-user SaaS

M2+ wraps the pipeline in Next.js 16 + FastAPI + Clerk with credits, SSE progress streaming and an async worker for concurrent runs.

Stack

Python 3.12GPT-5.4-miniClaude Sonnet 4.6Claude Haiku 4.5FastAPINext.js 16ClerkTavilySSESQLiteLaTeX
Next case study
FedFV-CV
Federated Deep Learning for Biometric Auth