Build a self-hosted open-weight AI chatbot platform on a cloud VM

We need an experienced AI/ML engineer to deploy a locally hosted large language model on a single cloud VM, make it easy to interact with, and lay the foundation for a multi-agent, multi-model system. The first working proof of concept (PoC) is due by the end of week 4. Background Our aim is locally hosted AI agents that work together, using more than one model, for business uses such as marketing, financial analysis or internal accounting (examples only). We prioritise a balance of cost and performance. Requests may be queued and heavy tasks run overnight in exchange for a fixed, predictable cost. Over time the platform must scale with hardware, gain skills such as web search and document analysis, and allow us to monitor and improve answer quality. Roadmap (12 weeks, weekly milestones) Week-01: Comparison study and recommendation (models, hosting options, VM size, cost) Week-02: Recommended model deployed on a single VM, with authenticated API Week-03: Chat interface, user accounts, basic monitoring, scheduled start/stop Week-04: PoC gate: benchmarks and two acceptance tests, demo Week-05: LangGraph agent foundation with tool calling Week-06: Document analysis skill (PDF, Word, Excel) with citations Week-07: Web search skill Week-08: Agents gate: multi-agent workflow demo, job queue and overnight scheduling Week-09: Multi-model routing and concurrency testing Week-10: Quality monitoring: test set, tracking, and recommendations for improving with retrieval-augmented generation (RAG) or fine-tuning Week-11: Containerisation and a documented, tested scale-up beyond one VM Week-12: Final gate: documentation, runbook, training, handover, 2-week fix period PoC success criteria (end of week 4): Generation speed of at least 30 tokens per second per request. Test 1, chatbot: answers questions and reasons through responses with acceptable response time. Report time to first token under normal use. Test 2, spreadsheet analysis: given a spreadsheet we supply, answers questions and analyses the data correctly. Pass mark: 90% of 20 questions. Test conditions stated (prompt length, number of users, number of runs), with median and worst-case results. Stable with several simultaneous users. Technical requirements Open-weight models are acceptable, but the licence must allow commercial use. LangGraph for agent orchestration. Containers and infrastructure as code, with no lock-in to one cloud provider. An OpenAI-compatible API and a simple web chat interface. Deliverables Code repository, documentation, benchmark scripts and results, demos at each gate, a runbook, and a handover and training session. Terms Confidentiality agreement required. All code and work product are owned by us. Work is done in our cloud accounts with least-privilege access. Approval is needed before provisioning resources. No real or confidential data in testing. We accept each gate by running the benchmarks and tests ourselves. Out of scope Outbound content filtering (a later phase). Scraping of platforms (a later phase).