Full-stack AI developer to set up a secure local GPT-like assistant with doc analysis and memory.

Seeking a highly competent AI engineer to build a fully local, modular AI assistant designed for legal analysis, document synthesis, clinical planning, and research-intensive writing. This is a mission-critical cognitive environment, not a chatbot or LLM toy. The system must run entirely offline, preserve compartmentalized memory, and support vector-based document ingestion — all deployed on a MacBook Pro M4 Max (64GB RAM, 2TB SSD). It must be auditable, inspectable, and fully under user control at every layer. Build Scope (Three Phases – One Builder Only) You will be contracted to deliver the entire architecture across three phases: Phase 1 — ("Legal & Tax Office")

• LM Studio / llama.cpp optimized for Apple Silicon
• Vector DB setup (Chroma/Qdrant/Weaviate) with local document parsing
• UI (Streamlit, Electron, or lightweight shell)
• PDF/DOCX ingestion with embedded citation + source trace
• Memory persistence: prompt saving, tone tagging, output control
• Output styles: affidavit, brief, memo, etc.
• No cloud sync. No logging. Fully local.

Phase 2 — ("The Board Room") Multi-Domain Strategic Workspace

• Add support for clinical workflows, estate planning, strategic writing
• Improved memory containers (per domain)
• API access wrapper to GPT-4 or Claude 3 (via user-owned keys ONLY)
• Enhanced persona switching, tone locks, exportable sessions

Phase 3 — ("The Den") Private Cognitive Module

• Separate vault for high-context reasoning and longform writing
• Firewalled from Office; no cross-contamination
• Persistent memory, stylistic voice preservation, session archiving
• Local-only inference with optional narrative assistive tools
• Optional future integration: multimodal parsing, encryption layers

Security & Audit Clauses (Non-Negotiable)

• You must consent to third-party code audit after each phase
• No proprietary wrappers
• No API calls unless explicitly declared and authorized
• All configuration paths, prompts, vector stores, and memory logs must be accessible to the user
• You will provide:
• Full install scripts
• Summary.txt documenting build logic, parameters, dependencies
• Notes to replicate install from clean OS

Payment Structure This is an hourly project with a firm "not-to-exceed" ceiling, payable at the end of each phase upon satisfactory completion and code audit. Terms:

• You (the builder) are committing to the entire 3-phase build, unless mutually terminated.
• I (the client) commit to paying for each completed phase promptly and fully once:
• Deliverables are met
• The audit review is cleared
• I reserve the right to end the project after any phase, for any reason, without obligation to proceed further.
• You agree that no partial or hidden ownership exists — all build components, logic, and configuration must be shared in full and are under the client’s control.

Summary: PhaseDeliverable Payment Exit Clause Phase 1 Inference + UI Paid after audit Client may exit Phase 2 Vector Store + API logic Paid after audit Client may exit Phase 3 Final Memory Shell + Vault Split Paid after audit Final Please submit your hourly rate and your not-to-exceed cap (USD) for the full build. System Context You’ll be working on a clean Apple M4 Max MacBook Pro (64GB RAM, 2TB SSD), macOS Sonoma. Pre-installed:

• LM Studio
• Homebrew
• Git

Available access:

• AnyDesk (if needed)
• Local data files (legal PDFs, notes, DOCXs, etc.)

Preferred Stack (flexible with reasoned alternatives) Layer Preferred Tool Inference LM Studio / llama.cpp Vector DB Chroma or Qdrant File Parsing PyMuPDF, unstructured.io UI Streamlit or Electron Memory Logic JSON store or LangGraph API Gateway (Phase 2–3) GPT-4/Claude via secure keys Models to install/test:

• MythoMax 13B Q5
• Chronos Hermes 13B
• Mistral 7B Instruct
(Final quantization/tuning TBD based on your advice)

✅ You Should Apply If You:

• Have built secure local GPT-style systems before
• Know vector stores, LLM inference, and UI wiring well
• Understand data compartmentalization
• Are comfortable being audited
• Can document your work cleanly
• Are discreet and professional in high-trust builds

Bonus if you’ve worked with:

• Apple Silicon
• Clinical/legal data environments
• High-context cognitive tools (not just chat interfaces)

How to Apply

In your message, please include: 1. A short list of similar builds (esp. local/private GPT) 2. Your preferred stack and why 3. Your hourly rate 4. Your not-to-exceed ceiling for all three phases 5. Confirmation you accept the audit clause and early termination terms Apply Job! Apply to this Job