INITIALIZING — PROSPECT ENGINE

000

PROSPECT · SYSTEM / LIVE — IN ACTIVE DEVELOPMENT

It reads the morning papers before I wake up.

Prospect is an autonomous job-prospecting engine: a multi-agent system that discovers job postings, matches them against persona-scoped resumes, and delivers ranked alerts — end to end, without me touching it.

11 STAGES·15–20 GB / DAY·RUNNING LOCAL

Why it exists

Job hunting is a pipeline problem. Postings are scattered across newspapers, boards, and feeds; matching them against a resume is retrieval; tailoring an application is generation. So I built it as a pipeline: OCR at the front, RAG in the middle, Telegram at the end.

Architecture

SCANOCR engineEXTRACTLLM pipelineEMBEDFAISS vectorsMATCHSemantic searchGENERATEResume tailoringDELIVERTelegram alertssmart-job-scanner-v2merlin-cli/bridgepersona-context-enginejob-discovery-enginemerlin-cli/bridgejobboard-api

Five components. Each is honest about its state.

smart-job-scanner-v2

11-stage OCR + LLM extraction pipeline; chews through 15–20 GB of newspaper PDFs daily; semantic matching; Telegram delivery

RUNNING LOCALGitHub →

merlin-cli / bridge

Tool-calling runtime executing sandboxed local commands from LLM decisions; generation engine

RUNNING LOCALPRIVATE

persona-context-engine

FAISS embeddings strict-mapping projects to three base resumes — no context bleeding between personas

RESEARCHPRIVATE

job-discovery-engine

LangGraph state machine orchestrating discovery; NVIDIA NIM integration

IN DEVELOPMENTPRIVATE

jobboard-api

Django REST + Postgres + Redis application tracker

SHIPPEDGitHub →

SHIPPED — running and done.

RUNNING LOCAL — battle-tested on my hardware, not published.

IN DEVELOPMENT — being built now.

RESEARCH — proven in experiments, not integrated.

Private components are described here and available on request.

Proof

SCREENSHOT — Sid to capture: Telegram alert arriving with ranked matches

SCREENSHOT — Sid to capture: pipeline run — pages OCR'd, blocks extracted

What I’d tell another engineer

Parts of this system are boring on purpose — the Telegram bot and the multi-key Gemini client are legacy code transplanted from v1 because they never failed. Parts are hard — stealth crawling and LaTeX-safe generation are still in development, and I say so above. A system that reports its own state honestly is the point.

Want this kind of pipeline thinking on your team?