AI Project Phase 1: Audit, Integrate, Harden and Hand Over an Offline Consulting Copilot
Budget: €250 – €750 EUR
Windows 11 | OpenClaw + LM Studio + Qdrant + Meeting Ingestion Pipeline
Project Overview
I am looking for a highly capable freelancer to work remotely via Windows Remote Desktop (RDP) on a dedicated Windows 11 Pro workstation and take over an existing partially implemented local AI stack.
This is not primarily a greenfield build. A partial implementation already exists and must first be audited, clarified, and brought into a clean, accepted specification before further implementation proceeds.
The target system is an offline, evidence-first consulting copilot that ingests meeting recordings and turns them into a searchable, citeable knowledge base. It must answer only from local evidence and must refuse or clearly mark uncertainty when evidence is insufficient.
This evidence-first behavior, reboot-safe operation, offline-first processing, and local-only architecture are already reflected in the existing project documentation and acceptance materials.
Important Working Principle
This project will be awarded in multiple sequential phases.
Each phase is a separate commission.
The next phase will only be awarded after the previous phase is accepted.
The purpose of this structure is to avoid ambiguity, prevent scope disputes, and ensure that implementation follows an agreed architecture and acceptance model.
Existing Baseline
A partial local stack already exists and should be audited and improved, not rewritten by default.
Existing or documented components include:
LM Studio serving a local model through an OpenAI-compatible API
Qdrant as local vector database
KB backend and UI
OpenClaw as local gateway and orchestration layer
Recording ingestion pipeline for transcription, diarization, chunking, indexing, and artifact extraction
Windows scripts and runbooks for startup, diagnostics, and reboot checks
Existing evidence-first prompting and refusal behavior concepts
A documented local architecture where OpenClaw should call the KB backend, and the KB backend handles retrieval and citations against LM Studio and Qdrant
Current documented default local ports:
LM Studio: http://localhost:1234/v1
Qdrant: http://localhost:6333
KB Backend/UI: http://localhost:8000
The existing codebase is documented as containing backend, ingestion, file watcher, speaker mapping, logs, scripts, docs, and environment-driven settings.
Scope Structure: Sequential Phases
Phase 1 - Audit, Specification Refinement, and Acceptance Criteria Definition
This phase is mandatory and comes first. No implementation work starts before this phase is accepted.
Objective
Review the current stack
Compare it against the intended outcome
Identify gaps and risks
Produce a corrected, detailed, implementation-ready specification
Produce clear acceptance criteria
Tasks
Audit the existing codebase, scripts, configuration, and runtime setup
Review the current architecture and determine what is:
working
partially working
unstable
missing
incorrectly designed
unclear in scope
Verify how the current flow works across:
OpenClaw
KB backend and UI
LM Studio
Qdrant
ingestion pipeline
Windows startup and reboot behavior
Review current evidence-first behavior and determine whether it is prompt-only or also enforced in backend logic
Review current ingestion outputs and citation traceability
Identify where the architecture is sound and should be preserved
Identify where selective refactoring is better than patching
Identify any components that would justify replacement only if necessary
Produce a refined project specification
Produce a phase-by-phase implementation plan
Produce revised acceptance criteria
Define what is in scope and out of scope for later phases
Deliverables
Short audit report
Recommended target architecture
Revised specification document
Revised acceptance criteria document
Implementation plan with risks, assumptions, and phase boundaries
Recommendation on whether any component should be rebuilt rather than repaired
Acceptance for Phase 1
Phase 1 is accepted if I receive a written package that allows me to answer these questions clearly:
What already works?
What is missing?
What must be fixed?
What should remain unchanged?
What are the acceptance criteria for each later phase?
Is a rewrite needed anywhere, and if so, why?
Technical Constraints
Offline-first: no external API calls for normal processing
Security: no client data leaves the machine
Evidence-first: factual claims must be supported by local evidence
Refusal behavior: unsupported questions must refuse or clearly mark uncertainty
Windows stability: reboot-safe, fixed ports, no duplicate instances
Documentation: all installs and changes must be scriptable and documented
Environment
Windows 11 Pro
RTX 5090
Intel Core Ultra 9
64 GB RAM
RDP access only
Preferred Technical Baseline
The current documented baseline includes:
LM Studio as local LLM server on localhost:1234
Qdrant on localhost:6333
KB backend/UI on localhost:8000
Existing model documentation around Qwen2.5-32B-Instruct Q4_K_M as a strong German-capable local model for this workstation class
You may propose changes, but only with clear justification.
Proposal Requirements
Please send:
Relevant experience and Git repository link
Offline and local LLM systems
LM Studio, llama.cpp, or OpenAI-compatible local APIs
Windows automation and service management
Qdrant or similar vector databases
Transcription and diarization pipelines
Secure and offline deployments
Phase 1 approach
How you would audit the current stack
What you would inspect first
What risks you expect
Architecture opinion
What should likely be preserved
What is likely to need refactoring
Under what conditions you would recommend replacement of any component
Delivery model
Estimated time and price for Phase 1 only
Optional estimate ranges for later phases
Milestone and payment structure
Software list
What you expect to install or modify
One-time downloads likely required
Any known offline limitations
Important Note to Applicants
Do not propose a full rewrite by default.
This project should begin with audit, clarification, and acceptance design.
A rewrite of any component is only acceptable if justified in Phase 1 with a clear technical and commercial rationale.
Project Overview
I am looking for a highly capable freelancer to work remotely via Windows Remote Desktop (RDP) on a dedicated Windows 11 Pro workstation and take over an existing partially implemented local AI stack.
This is not primarily a greenfield build. A partial implementation already exists and must first be audited, clarified, and brought into a clean, accepted specification before further implementation proceeds.
The target system is an offline, evidence-first consulting copilot that ingests meeting recordings and turns them into a searchable, citeable knowledge base. It must answer only from local evidence and must refuse or clearly mark uncertainty when evidence is insufficient.
This evidence-first behavior, reboot-safe operation, offline-first processing, and local-only architecture are already reflected in the existing project documentation and acceptance materials.
Important Working Principle
This project will be awarded in multiple sequential phases.
Each phase is a separate commission.
The next phase will only be awarded after the previous phase is accepted.
The purpose of this structure is to avoid ambiguity, prevent scope disputes, and ensure that implementation follows an agreed architecture and acceptance model.
Existing Baseline
A partial local stack already exists and should be audited and improved, not rewritten by default.
Existing or documented components include:
LM Studio serving a local model through an OpenAI-compatible API
Qdrant as local vector database
KB backend and UI
OpenClaw as local gateway and orchestration layer
Recording ingestion pipeline for transcription, diarization, chunking, indexing, and artifact extraction
Windows scripts and runbooks for startup, diagnostics, and reboot checks
Existing evidence-first prompting and refusal behavior concepts
A documented local architecture where OpenClaw should call the KB backend, and the KB backend handles retrieval and citations against LM Studio and Qdrant
Current documented default local ports:
LM Studio: http://localhost:1234/v1
Qdrant: http://localhost:6333
KB Backend/UI: http://localhost:8000
The existing codebase is documented as containing backend, ingestion, file watcher, speaker mapping, logs, scripts, docs, and environment-driven settings.
Scope Structure: Sequential Phases
Phase 1 - Audit, Specification Refinement, and Acceptance Criteria Definition
This phase is mandatory and comes first. No implementation work starts before this phase is accepted.
Objective
Review the current stack
Compare it against the intended outcome
Identify gaps and risks
Produce a corrected, detailed, implementation-ready specification
Produce clear acceptance criteria
Tasks
Audit the existing codebase, scripts, configuration, and runtime setup
Review the current architecture and determine what is:
working
partially working
unstable
missing
incorrectly designed
unclear in scope
Verify how the current flow works across:
OpenClaw
KB backend and UI
LM Studio
Qdrant
ingestion pipeline
Windows startup and reboot behavior
Review current evidence-first behavior and determine whether it is prompt-only or also enforced in backend logic
Review current ingestion outputs and citation traceability
Identify where the architecture is sound and should be preserved
Identify where selective refactoring is better than patching
Identify any components that would justify replacement only if necessary
Produce a refined project specification
Produce a phase-by-phase implementation plan
Produce revised acceptance criteria
Define what is in scope and out of scope for later phases
Deliverables
Short audit report
Recommended target architecture
Revised specification document
Revised acceptance criteria document
Implementation plan with risks, assumptions, and phase boundaries
Recommendation on whether any component should be rebuilt rather than repaired
Acceptance for Phase 1
Phase 1 is accepted if I receive a written package that allows me to answer these questions clearly:
What already works?
What is missing?
What must be fixed?
What should remain unchanged?
What are the acceptance criteria for each later phase?
Is a rewrite needed anywhere, and if so, why?
Technical Constraints
Offline-first: no external API calls for normal processing
Security: no client data leaves the machine
Evidence-first: factual claims must be supported by local evidence
Refusal behavior: unsupported questions must refuse or clearly mark uncertainty
Windows stability: reboot-safe, fixed ports, no duplicate instances
Documentation: all installs and changes must be scriptable and documented
Environment
Windows 11 Pro
RTX 5090
Intel Core Ultra 9
64 GB RAM
RDP access only
Preferred Technical Baseline
The current documented baseline includes:
LM Studio as local LLM server on localhost:1234
Qdrant on localhost:6333
KB backend/UI on localhost:8000
Existing model documentation around Qwen2.5-32B-Instruct Q4_K_M as a strong German-capable local model for this workstation class
You may propose changes, but only with clear justification.
Proposal Requirements
Please send:
Relevant experience and Git repository link
Offline and local LLM systems
LM Studio, llama.cpp, or OpenAI-compatible local APIs
Windows automation and service management
Qdrant or similar vector databases
Transcription and diarization pipelines
Secure and offline deployments
Phase 1 approach
How you would audit the current stack
What you would inspect first
What risks you expect
Architecture opinion
What should likely be preserved
What is likely to need refactoring
Under what conditions you would recommend replacement of any component
Delivery model
Estimated time and price for Phase 1 only
Optional estimate ranges for later phases
Milestone and payment structure
Software list
What you expect to install or modify
One-time downloads likely required
Any known offline limitations
Important Note to Applicants
Do not propose a full rewrite by default.
This project should begin with audit, clarification, and acceptance design.
A rewrite of any component is only acceptable if justified in Phase 1 with a clear technical and commercial rationale.