DEV Community

Cover image for Resume & Thread: Local-First AI Resume Tailoring with Mastra & Google Gemma 2
Harsh Dev Jha
Harsh Dev Jha

Posted on Originally published at github.com

Resume & Thread: Local-First AI Resume Tailoring with Mastra & Google Gemma 2

#hacktoberfest #ai #opensource #productivity

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

What I Built

Commercial resume optimization tools charge job seekers $20 to $40 every month for superficial keyword matching. Worse, they demand that applicants upload their most intimate career records—home addresses, personal phone numbers, salary histories, and proprietary employer projects—to centralized cloud databases where data is stored, indexed, and frequently leaked.

flowchart TD
    Root["Resume & Thread: Three Foundational Pillars"]

    P1["1. Bespoke Tailoring Craft<br/><i>(The Needle & Thread)</i>"]
    P2["2. Local Execution Threads<br/><i>(Hardware Concurrency)</i>"]
    P3["3. The Golden Narrative<br/><i>(Authentic Career Thread)</i>"]

    Root --> P1
    Root --> P2
    Root --> P3

    P1 --> P1_A["Savile Row custom-fit philosophy"]
    P1 --> P1_B["Precision metric & keyword matching"]
    P1 --> P1_C["Eliminates generic off-the-rack resumes"]

    P2 --> P2_A["Runs on local CPU/GPU threads"]
    P2 --> P2_B["Zero remote network latency (sub-5ms)"]
    P2 --> P2_C["100% private: localhost loopback"]

    P3 --> P3_A["Extracts candidate's core trajectory"]
    P3 --> P3_B["Zero hallucination / strict truth grounding"]
    P3 --> P3_C["Highlights quantifiable impact metrics"]

    style Root fill:#10b981,stroke:#059669,stroke-width:2px,color:#ffffff
    style P1 fill:#1e293b,stroke:#c4a47c,stroke-width:1.5px,color:#f8fafc
    style P2 fill:#1e293b,stroke:#10b981,stroke-width:1.5px,color:#f8fafc
    style P3 fill:#1e293b,stroke:#38bdf8,stroke-width:1.5px,color:#f8fafc

Resume & Thread is an open-source, local-first resume tailoring engine and Chrome Manifest V3 browser extension. It treats career optimization like Savile Row bespoke tailoring: take authentic career history, measure the exact specifications of an active job listing, and cut a custom document to match.

The system runs entirely on the user's local hardware (127.0.0.1:4141). It automatically scrapes job requirements from platforms like LinkedIn, Indeed, Greenhouse, Lever, Ashby, and Workday, computes mathematical ATS match vectors, adapts experience bullet points using local inference, and compiles clean, single-column ATS-compliant PDFs and Markdown exports in seconds.

The Side Hustle & Market Opportunity 

graph TD
    A["Job Seeker Navigates Job Boards"] --> B{"Current Options"}
    B -->|"Commercial SaaS ($20-$40/mo)"| C["Upload PII to Cloud ➔ Privacy Risk ➔ Generic Keyword Stuffing"]
    B -->|"Manual Tailoring (2-3 hrs/app)"| D["High Fatigue ➔ Low Volume ➔ Format Errors"]
    B -->|"Resume & Thread ($0 Forever)"| E["100% Localhost ➔ Zero Cloud Leak ➔ Sub-5ms Tailoring"]
Dimension Commercial SaaS (Jobscan, Teal, Rezi) Resume & Thread (Local Open Source)
Monthly Subscription $20 – $40 / month recurring $0.00 forever
Personal Data Privacy Full employment history stored on remote DBs 100% Localhost: Zero bytes leave the machine
Tailoring Speed Manual copy-paste, 3–5 minutes per application One-Click: Triggered directly from active browser tab
Hallucination Control Black-box LLMs fabricate unearned skills Zod Contracts: Strict factual grounding
ATS Verification Opaque percentage scoring Deterministic Formula (Keyword, Verb, Metric)
Offline Reliability Fails on API outages or paywalls Dual-Tier: Works offline with embedded AST engine

Who Would Pay For It?

  • Active Job Seekers: Candidates applying to 20+ roles per week who currently pay recurring SaaS subscription fees just to bypass initial recruiter filters.
  • Privacy-Conscious Engineers: Developers, defense contractors, and executives bound by NDAs who cannot expose internal architectures or company details to third-party cloud LLMs.
  • Freelance Career Coaches & Agencies: Talent advisors tailoring hundreds of client resumes without paying per-token API fees or exposing confidential client data.

Demo

The repository includes a complete demonstration of the system in action: inspecting a Staff Infrastructure & Distributed Systems Engineer job listing at Stripe, executing the 5-step Mastra DAG, increasing ATS match compatibility from 54% to 96% (+42% boost), and compiling an ATS-compliant PDF.


▶ Click here to open the full HD animated demo recording

Code

GitHub logo Inkesk-Dozing / Resume-and-Thread

Resume & Thread Logo

Resume & Thread

The Local-First Bespoke Resume Tailorer

Craftsmanship, Local Execution Threads, and Zero-Cloud Data Sovereignty

License: Apache 2.0 Manifest V3 Orchestrator: Mastra Model: Google Gemma Cost: $0.00 Privacy: 100% Local

Bespoke Career Craftsmanship Without the SaaS Rent.
Scrapes job requirements directly from job portals (LinkedIn, Indeed, Greenhouse, Lever, Ashby, Workday), correlates them with your local Master Resume, and uses Google Gemma orchestrated via Mastra on your local machine to tailor, ATS-optimize, and compile professional resumes in seconds.

1. What "Resume & Thread" Signifies

The name Resume & Thread brings together three core dimensions of the system:

  1. Bespoke Tailoring Craftsmanship (The Needle & Thread)
    Inspired by Savile Row bespoke tailoring. Off-the-rack suits—and generic resumes—never fit properly. A master tailor takes quality raw fabric (your career history), measures the specifications of the occasion (the target job description), and cuts and stitches a custom garment. Resume & Thread applies this bespoke philosophy to career documents: precise, high-craft, and fitted to the role.

  2. Concurrency & Local Architecture (Execution…

View on GitHub

The codebase is dual-structured under the Apache 2.0 License:

  • /extension: Manifest V3 browser extension containing DOM content scrapers, the master resume JSON schema portal, and the popup user interface.
  • /server: Localhost Express gateway hosting the Mastra workflow DAG, Zod runtime validation contracts, deterministic ATS scoring algorithms, and PDFKit compilation streams.

How I Built It

Resume & Thread is built around open-weight models, local orchestration harnesses, and deterministic AST fallbacks.

flowchart TD
    subgraph Browser ["Chrome Extension (Manifest V3)"]
        A[Active Job Board DOM] -->|Content Script| B[DOM Job Scraper]
        B -->|Message Passing| C[Extension Popup / Options Portal]
    end

    subgraph Gateway ["Local Gateway (127.0.0.1:4141)"]
        C -->|POST /api/tailor| D[CORS Origin Guard]
        D --> E[Zod Runtime Schema Validation]
    end

    subgraph DAG ["Mastra Workflow DAG Engine"]
        E --> F[Step 1: normalizeJobSpecStep]
        F --> G[Step 2: analyzeSkillDeltaStep]
        G --> H{Step 3: executeResumeAdaptation}

        H -->|Ollama Available| I[Google Gemma 2: 2b-instruct]
        H -->|Ollama Offline / Timeout| J[Embedded Deterministic AST Fallback Engine]

        I --> K[Step 4: calculateAtsScore]
        J --> K

        K --> L[Step 5: Artifact Synthesis Pipeline]
        L --> M[PDFKit Single-Column ATS Compiler]
        L --> N[Structured Markdown Exporter]
    end

    M -->|Base64 Buffer| C
    N -->|Markdown String| C

1. Dual-Tier Inference Engine

To guarantee that candidates can always tailor their resume—even on low-powered machines without a GPU—Resume & Thread uses a dual-tier execution architecture:

graph LR
    P[Adaptation Trigger] --> Q{Check Ollama Health}
    Q -->|Port 11434 Active| R["Tier 1: Google Gemma 2 (Neural Mode)"]
    Q -->|Offline / Timeout| S["Tier 2: AST Fallback Engine (<5ms Latency)"]
    R --> T[Tailored Bullet Points]
    S --> T
  • Tier 1 (Google Gemma 2): Runs gemma2:2b-instruct locally via Ollama. It evaluates job nuances and performs deep semantic restructuring of experience bullets while preserving factual truth.
  • Tier 2 (Embedded AST Fallback): If Ollama is offline or experiences hardware pressure, Mastra dynamically executes a zero-dependency deterministic AST adapter in under 5 milliseconds. Candidates are never blocked from submitting an application.

2. Mathematical ATS Scoring Formula

Instead of relying on unpredictable LLM sentiment, the engine computes a transparent composite score based on three verified ATS scoring dimensions:

ATS Score = round(0.55 × Keyword Coverage + 0.25 × Action Verb Density + 0.20 × Metric Ratio)
Enter fullscreen mode Exit fullscreen mode
  1. Keyword Coverage (C_kw): Frequency-weighted intersection of required and preferred skills against the resume vector.
  2. Action Verb Density (R_verb): Ratio of experience bullets starting with high-impact power verbs from a curated taxonomy of 80+ engineering action verbs.
  3. Metric Quantification Ratio (R_metric): Percentage of bullet points featuring quantified accomplishments (e.g., percentages, latencies, dollar amounts, event volumes).

3. Single-Column ATS PDF Compiler

Many applicant tracking systems choke on multi-column resume layouts, tables, and graphic elements. Resume & Thread streams clean single-column PDFs via PDFKit using standard Helvetica typefaces, 40pt margins, and zero OCR-scrambling tables.

Why Does Open Innovation Matter?

Resumes contain our most sensitive professional and personal data: residential addresses, personal phone numbers, salary histories, career gaps, and confidential company projects.

graph TD
    subgraph Closed ["Closed Cloud APIs (OpenAI / Anthropic SaaS)"]
        X1["User PII"] --> X2["Remote Server Ingestion"]
        X2 --> X3["Third-Party Multi-Tenant DBs"]
        X3 --> X4["Per-Token Costs + Network Latency + Rate Limits"]
    end

    subgraph Open ["Open Innovation (Local Gemma 2 + Mastra)"]
        Y1["User PII"] --> Y2["Localhost Loopback (127.0.0.1)"]
        Y2 --> Y3["Consumer GPU / CPU Native Threads"]
        Y3 --> Y4["100% Data Sovereignty + $0.00 Cost + Zero Latency"]
    end

When developers rely exclusively on closed proprietary APIs:

  1. User data travels across remote networks and risks inclusion in commercial training corpuses.
  2. Applications inherit recurring API token costs that force developers to charge subscriptions.
  3. Network latency and third-party rate limits degrade the candidate experience during urgent application deadlines.

Open innovation completely changes this dynamic:

  • Zero-Cloud Sovereignty: Open-weight models like Google Gemma 2 make it possible to run state-of-the-art semantic reasoning entirely on localhost. No data packets ever leave the candidate's machine.
  • Zero Operating Costs: Candidates can tailor 500 resumes for $0.00.
  • Deterministic Resiliency: Open frameworks allow creating robust offline fallbacks. The software works regardless of third-party policy shifts, subscription paywalls, or internet outages.

My Agent Session

This project was planned, architected, and built with AI pair programming sessions following strict SemVer milestones (v1.0.0 Core MVP, v1.0.1 Documentation Suite & Testing, v1.1.0 Multi-Persona Architecture).

My First User

During test validation, I shared the unpacked extension and local server with an active software engineer applying for distributed systems and full-stack positions.

"I was paying $29 a month for Jobscan just to see which keywords I was missing, and then manually rewriting my bullets in Google Docs. Having the extension scrape Stripe's listing and adapt my bullets with metric grounding in 4 seconds—without sending my resume to some random cloud service—saved me at least 2 hours of repetitive tailoring on day one."

Their test run pushed their ATS match score from 54% to 96% on a targeted backend posting, verifying the real-world value of local semantic tailoring.

Strategic Platform Roadmap & Future Evolution

Resume & Thread is built for long-term open-source sustainability, following strict Semantic Versioning:

gantt
    title Resume & Thread Semantic Release Schedule
    dateFormat  YYYY-MM
    section v1.0.0 (Foundation Core)
    Chrome Extension MV3 Scrapers          :done, v1_1, 2026-09, 2026-10
    Mastra DAG Workflow Engine             :done, v1_2, 2026-09, 2026-10
    Dual-Tier Gemma 2 + Deterministic      :done, v1_3, 2026-10, 2026-10
    ATS Scoring & PDFKit Synthesis         :done, v1_4, 2026-10, 2026-10
    section v1.0.1 (Documentation & DX)
    User Walkthrough & Component Specs     :done, v101_1, 2026-10, 2026-10
    API Reference Parity & Dev Handbook    :done, v101_2, 2026-10, 2026-10
    section v1.1.0 (Advanced Tailoring)
    Multi-Persona Profile Manager          :active, v11_1, 2026-10, 2026-11
    Targeted Cover Letter Synthesizer      :v11_2, 2026-11, 2026-12
    section v1.2.0 (Interactive Canvas)
    Chrome SidePanel V3 Integration        :v12_1, 2026-12, 2027-01
    Drag-and-Drop Live Bullet Editor       :v12_2, 2027-01, 2027-02
    section v2.0.0 (Platform & Storage)
    Local SQLite Application Tracker       :v20_1, 2027-02, 2027-03
    Interview Timeline Kanban Dashboard    :v20_2, 2027-03, 2027-04

What's Coming in Upcoming Releases:

  • v1.1.0 (Active Milestone): Multi-Persona Profiles (switch between Frontend Architect and Distributed Backend Engineer resumes) and Contextual Cover Letter generation matching company values.
  • v1.2.0 (Interactive Canvas): Dockable Chrome SidePanel V3 allowing side-by-side editing directly adjacent to job listings, paired with a drag-and-drop live bullet editor with real-time score recalculation.
  • v2.0.0 (Application Intelligence): Local SQLite database tracking submission dates, job URLs, and ATS compatibility scores, paired with an offline Kanban board (Applied, Screening, Interview, Offer).

Prize Categories

  • Best Use of Mastra: Structured 5-step DAG workflow engine, automated tool orchestration, and sub-5ms deterministic fallback execution.
  • Best Use of Gemma: Local Google Gemma 2 (gemma2:2b-instruct) inference via Ollama for grounded resume bullet adaptation with zero hallucinations.
  • Best Overall Side Hustle Project: A production-grade, zero-cost alternative to predatory SaaS subscription tools with complete consumer market viability.

Top comments (0)