AI Senior Lead Developer
At PVY.swiss LTD we are bunch of steam punk rabbits, who like to think ahead and contribute our ideas and learnings from each other and trough AI, into product improvements. We are focussing on open source and open weight models, but also include commercial AI API Endpoint with value added Services for our Customers such as from Anthropic, offering a unified hybrid SSE API Stream with innovative features.
Objective
- Continuation of Architecting, refine, and maintain the unified Server-Sent Events (SSE) API that powers real-time streaming across all PVYai endpoints
- Optimize the SSE stream pipeline, including
stream_wrapper,StreamingResponsewithtext/event-streamcontent type, and analytics capture from final SSE chunks - Ensure robust handling of streaming vs. non-streaming fallback paths for API clients without
chat_id - Collaborate on analytics integration, ensuring usage data from API-only sessions is captured and written to the analytics dashboard via sentinel chat mechanisms
- Maintain SSE injection capabilities for pre-flight operations (e.g., wing selection, tool discovery) directly into the SSE stream
Job Description
We are seeking an AI Senior Lead Developer to drive the evolution of our AI platform at PVY.swiss. You will be the technical authority behind our unified SSE API, LLM post-training pipelines, agentic loop designs, and customer-specific model training programs.
You will work hand-in-hand with our CTO and the Software development team to refine the PVYai Agent and shape upcoming PVY applications across our product ecosystem — including PVYvault, PVYoffice, PVYfiles, PVYbusiness OS, PVYserviceDesk, PVYagentic to name a few, and constantly improving our unified SSE API Stream.
This is a senior, hands-on leadership role requiring deep expertise in LLM engineering, real-time streaming APIs, MLOps, and agentic system design, training Models and self-improvement loop design.
Responsibilities
LLM Post-Training & Fine-Tuning
- Design and execute post-training pipelines for large language models, including SFT (Supervised Fine-Tuning), LoRA/QLoRA adapters, and preference optimization (DPO/RLHF)
- Manage model lifecycle on GPU infrastructure (A100 80GB, H200) using vLLM with FP8 quantization, Flash Attention, and long-context support (131K+ context windows)
- Optimize model deployment for co-resident scenarios (e.g., vLLM + Ollama sharing a single GPU with 92GB free VRAM for 3–4 models)
- Evaluate and benchmark post-trained models for quality, latency, and resource efficiency
Agentic Loop Design
- Design and refine agentic loop architectures for the PVYai Agent, including the 2-step Tool Discovery Pattern (
discover_mcp_tools→call_mcp_tool) that achieves 99.6% context reduction by serving only 2 meta-tool schemas instead of 252 - Engineer loop control flow for multi-model environments (Claude Opus 4/5, DSV4, GLM, Kimi3, etc.)
- Refinements for sentinel-based safety mechanisms: destructive prevention (bash-only interception, commit message stripping, regex refinement), Toolwise, SWE/Terminal validation warnings
- Design synthetic tool call + tool result patterns for validation injection (per NeMo Guardrails research)
- Optimize loop latency, prevent thinking loops, and ensure graceful degradation, evicting and re-onboarding
Training LLMs on Customer Data
- Build secure, PriVacY-preserving customer data training pipelines with PII detection and bypass controls (e.g., PVY PriVacY PII bypass for authorized presets)
- Develop data preparation, augmentation, and validation workflows for customer-specific fine-tuning
- Ensure data isolation and confidentiality across customer tenants (anyway given)
- Create evaluation frameworks to measure domain-specific performance gains from customer data training
- Establish reproducible training runs with versioned datasets, model artifacts, and rollback capabilities
PVYai Agent & Upcoming Application Development
- Work closely with our CTO and Software Developers to refine the PVYai Agent — including system prompt design, MCP server integration and dual-protocol support (OpenAPI/mcpo + MCP Streamable HTTP)
- Co-develop upcoming PVY applications that integrate AI capabilities across the product suite
- Implement and maintain the PVYMCPManager dual-protocol cache (TTL=3600s, 13 servers, 252 tools)
- Constantly refine PVYai Memory, PVYai Sentinel and its features
- Contribute to architecture decision records (ADRs) and technical documentation
Required Skills
- 5+ years of professional experience in AI/ML engineering, with at least 2 years in a senior or lead role
- Deep expertise in LLM training and post-training techniques (SFT, LoRA, QLoRA, DPO, RLHF)
- Strong experience with vLLM, Ollama, or equivalent inference engines, including quantization (FP8, INT4, AWQ) and Flash Attention
- Production experience with Server-Sent Events (SSE) streaming APIs in Python (FastAPI,
StreamingResponse, async generators) - Hands-on experience with agentic AI frameworks and tool-calling patterns (function calling, MCP protocol)
- Proficiency in Python (advanced), with experience in async/await, aiohttp, and real-time streaming
- Experience with GPU infrastructure management (A100/H200, VRAM budgeting, co-resident model deployment)
- Familiarity with Docker, systemd, and container orchestration for ML workloads
- Experience with PriVacY-preserving ML (PII detection, data anonymization, tenant isolation)
- Strong understanding of MLOps practices: model versioning, CI/CD for ML, evaluation pipelines, monitoring
Preferred Skills & Qualifications
- Experience with MCP (Model Context Protocol) servers and the OpenAPI/mcpo bridge pattern
- Knowledge of ChromaDB or equivalent vector databases for RAG pipelines
- Experience with Open WebUI or similar LLM frontend platforms
- Familiarity with NeMo Guardrails or equivalent safety/validation frameworks
- Contributions to open-source AI/ML projects
- Experience in a B2B SaaS or enterprise AI product environment
- Swiss residency or EU work eligibility
Technical Environment
| Component | Technology |
|---|---|
| Inference Engine | vLLM (FP16, FP8, H200/B300), OpenWeb-UI (modified) |
| API Framework | FastAPI, SSE streaming, OpenAI-compatible API |
| Agent Architecture | MCP dual-protocol, Tool Discovery Pattern, PVYai Sentinel |
| Model Serving | Claude Opus 4 (API), DeepSeek R1 32B (GPU), Qwen 3.5 27B (GPU) |
| Vector Store | ChromaDB (RAG embeddings) |
| Infrastructure | Docker, systemd,NVIDIA GPUs, ComfyUI, LoRa, Flux, Turbo etc |
| MCP Servers | 13 servers, 252+ tools (Filesystem, GitHub, GitLab, Docker, SQLite, Playwright, Draw.io, PDF2MD, Memory, Webfetch, Time, Git Local) |
What We Offer
- Modern Lifestyle and Working Environment. Top modern IT and Data Center Infrastructure, owned to the last screw. Industry leading Partners & Board Members
- Company issued Notebook / Power Stations with Accessoires, Home Lab Cluster Nodes or DC Cluster Node
- International oriented Team, regular Team Events within Switzerland, EU and Non-EU Cities.
Benefits
- Progressive and contemporary remuneration
- Participation on Companies Stock Program for Employees