Syed Mahsam

Full-Stack AI Developer

LaunchBoom • Freiberufler

May 2023 - Present • 3 yrs 5 mos

- Personally architect and hands-on build LaunchBoom's real-time campaign analytics engine in Python (Django REST), owning the data layer behind 500+ crowdfunding campaigns and $175M+ in funds raised still writing production code daily. - Built the campaign manager dashboard frontend using React and TypeScript, rendering live funding metrics and conversion analytics with real-time updates, cutting manual reporting cycles by over 60%. - Design and lead development of multiple AI agents in Python/LangChain including an agentic copilot that autonomously analyzes historical campaign performance and backer demographics to generate launch playbooks, cutting pre-launch planning from weeks to hours. - Architected LaunchBoom's RAG knowledge-assistant platform end-to-end embedding pipeline, chunking strategy, and Pinecone vector-retrieval layer via LangChain. - Own the full LLM fine-tuning pipeline for custom LLaMA 3 and Mistral models (LoRA/QLoRA on AWS SageMaker), achieving a 32% accuracy gain at ~4x lower inference cost versus GPT-4. - Integrated and customized CMS workflows for campaign-page content management, streamlining how creators publish and update live campaign content. - Leverage AI-assisted development tools (e.g., Claude, GitHub Copilot) throughout the development lifecycle from code generation to debugging and to accelerate delivery of new features without compromising code quality. - Containerized and deployed backend services with Docker and orchestrated CI/CD pipelines for the analytics and AI services, enabling reliable, repeatable releases across environments.

Senior Full-Stack Engineer

Dealership Performance CRM • Freiberufler

Feb 2018 - Feb 2023 • 5 yrs

Architected FastAPI and Litestar AI inference microservices with PydanticAI schema validation, achieving 99.9% uptime across 250,000 daily transactions while isolating LLM workloads from the core application layer. Designed and deployed multi-stage RAG pipelines using LangChain, LlamaIndex, and HuggingFace sentence-transformers combining recursive text splitting, metadata-aware chunking, and hybrid BM25 + dense vector retrieval with MMR reranking, improving AI response relevance by 35%. Fine-tuned transformer models using PyTorch, TensorFlow, Keras, and HuggingFace PEFT (LoRA/QLoRA) on domain-specific CRM data, improving semantic retrieval accuracy over base models by 35% Engineered multi-agent orchestration systems using CrewAI and DSPyCRM autonomous agents queried external APIs, summarized call transcripts, and wrote typed outputs to PostgreSQL via SQLAlchemy, reducing manual data enrichment overhead by 40% Orchestrated AI data pipelines using Apache Airflow and Dagster for scheduled embedding generation, model retraining triggers, and vector index refresh, ensuring consistent RAG retrieval context across all production endpoints. Processed and transformed large-scale CRM datasets using Pandas, Polars, and NumPy for feature engineering and lead scoring, delivering business intelligence reports via Plotly and Seaborn visualizations. Architected Celery async task queues with Redis broker for non-blocking LLM inference and embedding generation, sustaining sub-100ms API response times under peak concurrent load Instrumented LLM pipelines with LangSmith and Langfuse for end-to-end observability — monitoring token usage, latency, hallucination rate, and retrieval hit rate, reducing cost-per-query by 22% Implemented streaming inference pipelines in FastAPI using SSE and WebSockets with backpressure handling for real-time token-by-token LLM output delivery. Led a cross-functional team of 5 engineers across Python AI, backend, and frontend driving a 25% sprint velocity