What I Built

Independent freelancers lose thousands of dollars every year to two expensive, recurring communication traps:

1. Unstated Brief Assumptions: Accepting client briefs that conceal unstated deliverables, missing dielines, undefined deadlines, and zero commercial clarity.
2. Unbudgeted Scope Creep: Receiving "just one quick little change" halfway through a project that quietly adds days of uncompensated labor.

Most freelancers cannot use cloud-hosted AI tools (like ChatGPT or Claude) to analyze their client contracts because doing so violates client Non-Disclosure Agreements (NDAs), exposes unreleased product IP, and leaks confidential commercial terms to third-party model providers.

SoloDesk is a private, local Windows desktop productivity workspace built for solo freelancers. It provides deterministic, AI-assisted boundary protection directly on consumer laptop hardware without sending a single byte of client data to the cloud:

• BriefCheck: Ingests unstructured client inquiries, extracts verified factual requirements, isolates unstated assumptions, flags critical omissions, and generates clarifying questions.
• ScopeShield: Semantically compares incoming client change requests against an approved, SQLite-locked scope version, producing itemized classifications (INCLUDED, POTENTIALLY_OUTSIDE, NEEDS_CLARIFICATION) with cited contract excerpts and editable, courteous draft replies.
• PolitePay: Multi-currency invoice tracking (USD, EUR, GBP, NGN, KES, CAD, AUD) with integer minor-unit math and calendar-date rules; generates polite, fact-grounded payment reminder drafts across gentle, polite, and firm tones without inventing unauthorized late fees.
• LeadLoop: Inbound lead qualification pipeline with currency-segregated unweighted totals and local follow-up drafting.
• Repeat Work: Retainer tracking and model-assisted renewal proposal drafting preventing fabricated acceptance claims.
• SoloDesk Boundary Policy: Explicit system invariant stating that SoloDesk never processes payments, sends messages, or provides legal advice. All outputs are strictly advisory drafts for human review.

Proposed Value & Willingness to Pay

SoloDesk addresses an immediate financial loss: absorbing a single out-of-scope feature or unbudgeted revision round costs a freelancer between $250 and $1,500 in unbilled time. Based on these economics, our commercial hypothesis is a $19/month or $149 lifetime local desktop license. (Note: This is an empirical pricing hypothesis based on freelancer pain points; per our anti-fabrication guidelines, no live commercial revenue has been collected prior to challenge release).

Demo

• Video Walkthrough:
  The walkthrough shows the Windows interface, client/project creation, BriefCheck running against the local model, ScopeShield comparing a request, and invoice/service-plan interactions. The figures and records shown are demonstration data, not evidence of business revenue.

• Windows Desktop Release: Download SoloDesk v0.1.0-challenge (Zip)

Prerequisites & Quick Start

SoloDesk runs locally on 64-bit Windows 10/11:

1. Ollama: Install Ollama for Windows and pull the open-weight model:

   ollama pull qwen3:1.7b

1. Download & Extract: Download solodesk-v0.1.0-windows-x64.zip from our GitHub releases.
2. Launch: Run launch-solodesk.bat (or start the Mastra service in service/ and run solodesk.exe).
3. Offline Validation: Disconnect Wi-Fi and execute BriefCheck or ScopeShield—all inference runs 100% locally.

Code

• GitHub Repository:

  joyy-dev-organization / solodesk

  SoloDesk

  SoloDesk is a private, local desktop productivity workspace designed for independent freelancers. It assists with brief analysis (BriefCheck), scope verification (ScopeShield), polite payment tracking (PolitePay), client pipeline tracking (LeadLoop), and follow-up proposal drafting (Repeat Work) using local open-weight models (qwen3:1.7b) and inspectable Mastra agent workflows, without sending client data to third-party cloud inference services.

  Note

  Release Status: v0.1.0-challenge · Feature Freeze Active (as of Sunday 2026-10-18T12:00 Africa/Lagos). All completed features and roadmap boundaries are locked for challenge submission.

  ---

  Submission Metadata & Challenge Traceability

  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  

  Attribute	Challenge Value
  Challenge	Hacktoberfest 2026 / DEV Challenge (Week 2: Best Use of Mastra)
  Prize Category	Best Use of Mastra (Orchestrating agents over local open models)
  Release Tag	v0.1.0-challenge
  Baseline Submission Commit	dd47789 (and M9 release packaging commits)
  Repository URL	https://github.com/joyy-dev-organization/solodesk
  Public DEV Post URL	[Pending Publication]
  Video Walkthrough Demo	[Pending Upload]

  …

  
  

  
  

  View on GitHub

  
  
• Release Tag: v0.1.0-challenge
• License: Permissive MIT License © 2026 ladydiamond

How I Built It

SoloDesk is architected as a clean two-tier local system separating high-performance native desktop rendering from deterministic agent workflow orchestration:

flowchart LR
    subgraph UI ["Native Flutter Windows Desktop"]
        direction TB
        AppShell["App Shell & Navigation"]
        Screens["BriefCheck · ScopeShield · PolitePay · LeadLoop · RepeatWork"]
        SQLite[("Local SQLite V4 DB\n%APPDATA%\\solodesk\\solodesk.db\n(Immutability Triggers)")]
        AppShell --> Screens
        Screens <--> SQLite
    end

    subgraph Service ["Local Loopback Service (127.0.0.1:3141)"]
        direction TB
        Auth["Bearer Token Auth (.runtime-token)\n+ CORS Origin Rejection"]
        Mutex["Atomic Concurrency Mutex\n(120s Timeout)"]
        Mastra["Mastra Deterministic Workflows\n(BriefCheck · ScopeShield · PolitePay · OfferDraft)"]
        Normalizer["Schema Normalizers\n(Zod Validation + Anti-Invention Filters)"]
        Auth --> Mutex --> Mastra --> Normalizer
    end

    subgraph Inference ["Local Ollama Runtime (127.0.0.1:11434)"]
        Ollama["Ollama Engine\nqwen3:1.7b (Open-Weight Model)\nOLLAMA_NO_CLOUD=1"]
    end

    Screens <== "HTTP Loopback (JSON)" ==> Auth
    Mastra <== "Loopback Model Stream" ==> Ollama

1. Flutter Windows Desktop Shell

Built with Flutter 3.44 and Dart 3.12, providing responsive desktop ergonomics, sidebar navigation, real-time elapsed request timers, dark mode slate theme, and direct SQLite integration (sqflite_common_ffi). Scope immutability is enforced at the database engine level via SQLite triggers that abort any UPDATE or DELETE operations on approved scopes.

2. Local Mastra Workflow Engine (@mastra/core 1.78.0)

The local backend is a standalone TypeScript service running strictly on loopback 127.0.0.1:3141. Instead of unconstrained prompt chaining, we orchestrate multi-step deterministic workflows:

• BriefCheckWorkflow (5 Steps): validate-input → select-niche-checklist (Design, Writing, General) → analyze-with-agent → validate-output → return-draft.
• ScopeShieldWorkflow (5 Steps): validate-scope-request → load-approved-context → compare-with-agent → validate-evidence-classification → return-editable-reply.
• InvoiceReminderWorkflow (4 Steps): validate-invoice-facts → prepare-reminder-prompt → generate-draft-with-agent → validate-and-normalize.
• OfferDraftWorkflow (4 Steps): validate-plan-facts → prepare-offer-prompt → generate-offer-draft → validate-and-normalize-offer.

3. Local Security Boundary

• Loopback Isolation: Binds strictly to 127.0.0.1:3141. External network interfaces are never opened.
• CSRF & DNS Rebinding Defense: Inspects incoming HTTP headers; browser Origin headers are rejected with HTTP 403 Forbidden.
• Ephemeral Bearer Authentication: The service generates a cryptographic .runtime-token upon startup; the Flutter client discovers this token via local process resolution.
• Single-Concurrency Mutex: Small local models running on CPU/shared RAM cannot handle parallel requests. An atomic mutex queues incoming requests and protects the local Ollama instance.

4. 12-Case Live Evaluation & Benchmark Telemetry

In Milestone 8, we evaluated SoloDesk against a pre-registered 12-case synthetic benchmark across Design, Writing, and General freelance niches (6 BriefCheck cases, 6 ScopeShield cases) running live against qwen3:1.7b:

• Schema Validation Rate: 100% (12/12) parsed cleanly against Zod schemas.
• Average 6-Point Human Rubric Score: 5.8 / 6.0 (assessing grounding, no missing fact fabrication, justified classifications, advisory replies, schema validity, and latency measurement).
• Cold Start Latency: 17.4s (cold model load).
• Warm Average Latency: 38.9s (range: 28.3s – 67.1s).
• Peak Process RSS Memory: 175.5 MB (safe for shared 16GB RAM laptop hardware).

Hardest Verified Bug & Solution

During M2 testing, our synthetic demo toggle became stuck on "Demo Data: ON". Inspection of %APPDATA%\solodesk\solodesk.db revealed that SQLite's CREATE TRIGGER IF NOT EXISTS is a no-op if the trigger already exists on disk. An earlier trigger lacking synthetic record bypass prevented purging synthetic records by throwing SQLite error 1811 (IMMUTABILITY_VIOLATION). We solved this by implementing _ensureTriggers inside SQLite's onOpen hook, automatically dropping and recreating triggers on database startup, and wrapping the multi-table purge inside an atomic db.transaction(...).

Why Does Open Innovation Matter?

Open innovation is not just an ideological preference—for solo freelancers, it is the only architecture that makes business sense:

1. Client Confidentiality & NDA Compliance: A freelance developer building a stealth startup MVP or a designer branding an unannounced product cannot paste client requirements into cloud-hosted LLM endpoints. Open-weight models (qwen3:1.7b) running on local open-source orchestration (Mastra) guarantee that sensitive trade secrets never leave the workstation.
2. Inspectable, Deterministic Orchestration: Cloud AI wrappers are notorious for silent model drifts and prompt breakages. Mastra lets us inspect every pipeline step, validate inputs and outputs against Zod schemas, and catch hallucinations before they reach the user.
3. True Vendor Independence & Offline Autonomy: Freelancers work from coffee shops, client sites, trains, and flights. SoloDesk works completely disconnected from the internet. There are no API rate limits, surprise token billing, or deprecation notices.

Honest Architectural Trade-Offs

• Inference Latency: Running quantized open models on consumer laptop CPUs (Intel UHD graphics) takes 25s to 60s per inference, compared to 2s on cloud APIs. We cushion this with responsive background polling, real-time elapsed timers, and cancelable requests.
• Small-Model Boundary Limits: 1.7B parameter models can struggle with subjective aesthetic nuances ("make it really pop"). We resolve this by enforcing defensive post-processing normalizers that convert ambiguous aesthetic requests into structured clarifying questions rather than ungrounded assumptions.
• Storage at Rest: Our local SQLite database is stored unencrypted in %APPDATA%\solodesk\ and relies on host OS user permissions.

Prize Categories

Best Use of Mastra

SoloDesk qualifies for Best Use of Mastra by demonstrating genuine, deeply integrated, and inspectable use of @mastra/core across all four operational business domains:

1. Deterministic Multi-Step Workflows: We reject monolithic prompt chaining in favor of discrete, typed Mastra steps (createWorkflow, createStep) that enforce input validation, domain checklists, and post-inference output normalization.
2. BriefCheckWorkflow: 5-step workflow utilizing specialized domain checklists (Design, Writing, General) and strict anti-invention constraints prohibiting fabricated prices or unstated deadlines.
3. ScopeShieldWorkflow: 5-step comparison workflow taking structured SQLite scope items and incoming client requests to produce itemized semantic classifications with text-anchored citations.
4. InvoiceReminderWorkflow & OfferDraftWorkflow: Multi-step drafting workflows enforcing strict arithmetic invariants (minor units, calendar dates) and preventing small models from inventing unauthorized collection penalties or claiming a client already accepted a renewal.
5. Local Open-Weight Model Integration: Mastra interfaces directly with local Ollama (qwen3:1.7b) over loopback, proving that sophisticated multi-agent orchestration does not require cloud LLM dependencies.