Hi, I'm Swapnil Desai 👋
Lead AI Engineer • Building & Scaling Production AI Agents • Multi-Agent Systems Architect
> Currently shipping AI systems that serve millions of customers across Europe. 12+ years in tech (8+ years Data Science, 3+ years Software Engineering). I also love teaching Data Science and helping developers build production-ready applications.
🚀 What I'm Doing Now
Lead AI Engineer at Deutsche Telekom Building, scaling, and governing AI Agents using Google's Agent Development Kit (ADK), Agent Builder, and Agent Engine.
🏆 Featured: OneBot (Multi-Agent Customer Support)
- Live in Production: Serving millions of customers at telekom.de
- Handling 7,500+ concurrent sessions and 25,000+ questions daily—reducing human agent escalations by 40% while driving 20% year-over-year operational savings
- Multi-agent architecture answering complex customer support queries across Telekom's European markets
- Built with enterprise-grade reliability and governance
🛠️ What I Build
I build end-to-end AI agents and ship production systems that can be used immediately—which solve real problems and add value to the org/customer.
- AI Agents: Single-agent → Multi-agent → MCP-based → Browser agents → Voice agents → Local agents
- Agent Architectures: Crews, swarms, hierarchical planning, tool orchestration, memory management, eval frameworks
- RAG Systems: Simple chains → Agentic RAG → Hybrid search (BM25 + Cross-encoders) → Local RAG with Gemma/Llama
- Chat-With-Anything: GitHub, Gmail, PDFs, videos, research papers (production-grade parsing + retrieval)
- Fine-Tuning: Gemma, Llama, and OSS models using PEFT/LoRA/QLoRA for instruction tuning
- Production Systems: Conversation AI platforms with human-in-the-loop, monitoring, and governance
💼 Tech Stack
AI & LLMs
GPT-4/GPT-4o Llama3 Gemma LangChain LangGraph ADK DSPy HuggingFace Transformers RAG Prompt Engineering
Multi-Agent & Tools
MCP Agent Development Kit CrewAI AutoGen OpenAI SDK Function Calling Tool Use Agent Evaluation
Data & Vector DBs
PySpark SQL PostgreSQL pgvector Qdrant Neo4j Redis Milvus NumPy Pandas
MLOps & Infrastructure
Docker Kubernetes FastAPI MLflow Airflow ArgoCD AWS (SageMaker, ECS, ECR, S3) Azure OpenAI GitLab CI/CD DVC Argilla
Frameworks & Languages
Python PyTorch Keras Streamlit FastAPI Linux Git
Production Load Testing & Optimization
Locust K6 JMeter
- Throughput Optimization: Designed systems handling 7,500+ concurrent sessions with optimized inference pipelines, batching strategies, and async processing to maximize requests/sec
- Latency Optimization: Sub-second responses via model quantization (GGUF/AWQ), efficient vector retrieval (HNSW), and Redis caching layers
- Load Testing at Scale: Distributed Locust clusters simulating 25,000+ daily queries to identify bottlenecks in agent orchestration and tool latency
- Production Profiling: GPU utilization tuning, memory leak detection, and auto-scaling strategies for peak traffic
🎓 Education & Experience
- M.Sc. in Computer Science, University of Pune (2013)
- 12+ years IT experience: 8+ years as Data Scientist, 3 years as Software Engineer
- Specialized in Conversational AI and making AI applications production-ready at scale
🌍 Community & Teaching
I'm passionate about education and building the AI engineering community:
- Teaching AI/Agents/Classical ML/DL across top EdTech platforms
- Helping students and professionals upskill with hands-on, production-focused curriculum
- Sharing real-world patterns for Agent Dev, Multi-Agent teams, RAG, and LLMOps
📫 Let's Connect
- 💼 Portfolio: [Your Portfolio Link]
- ✉️ Email: swapnil89.desai@gmail.com
- 💡 Medium: [Your Medium Articles]
- 🏢 LinkedIn: [Your LinkedIn]
- 🐦 Other: [Twitter/X or other relevant]
Open to collaborating on: AI Agents, Multi-Agent Systems, RAG, Conversational AI, MLOps, and Production LLM Applications.
⚡ Avid reader in AI, programming, and system design. Committed to continuous learning and shipping code that actually works in production.