Hi Reddit community! I’ve recently started my journey from being a full-stack developer working mainly with Laravel and the LAMP stack to a role focused on Generative AI. My primary responsibility involves integrating large language models using frameworks such as LangChain and Langraph, along with monitoring these models through LangSmith.
Currently, I’m working on implementing Retrieval-Augmented Generation (RAG) strategies using ChromaDB to address specific business challenges and minimize inaccuracies in responses. I’m still acquiring knowledge in this area.
Next on my learning agenda is mastering LangSmith for agent functionalities and fine-tuning models. Eventually, I aim to explore multi-modal use cases, including image processing.
After two months in this role, I find myself primarily engaged in web development, focusing on orchestrating LLM calls for smarter SaaS solutions. I am primarily using Django and FastAPI.
I hope to transition fully into a dedicated Generative AI role within the next 3-4 months. For those currently in Generative AI positions, I’m curious about your daily tasks. Do you encounter similar challenges, or is your experience quite different?
I would greatly appreciate any suggestions on topics to concentrate on and valuable resources that could assist me on this journey. Thank you for your support!
You’re already on the right track with RAG. I’ve been doing this for a couple years - the day-to-day changes a lot based on how mature your company is with AI.
All that orchestration work you’re doing? Perfect prep. Most AI roles are still heavy on integration problems, data pipelines, and making models actually work in production.
Here’s what I’d focus on:
Get solid with prompt engineering and evaluation frameworks. LangSmith’s great, but also try Phoenix or Weights & Biases for monitoring.
Learn vector databases beyond ChromaDB. I’ve used Pinecone, Weaviate, and Qdrant - they each shine in different situations.
Your 3-4 month timeline’s realistic if you keep building real projects. Market’s hot for people who can ship AI products, not just mess around with models.
Contributing to LangChain open source helped me a ton. You learn fast and build credibility.
Multi-modal stuff’s fun, but text-based RAG and agent workflows are where the money is right now. Nail those first.
Two months in and you’re handling LLM orchestration? That’s solid progress.
Here’s what everyone misses - those integration headaches get way easier with full pipeline automation. I’ve built AI workflows for years, and the biggest time sink is always connecting services, handling API calls, managing data flows between vector DBs and LLMs.
Stop manually coding every FastAPI endpoint and Django integration. Think bigger. Build automated workflows that handle your RAG pipeline end-to-end. Trigger data ingestion when new docs appear, auto-chunk and embed them, route queries to the right models, handle responses.
Last year I built a system that monitors our knowledge base, auto-updates ChromaDB embeddings, and routes complex queries through different LLM chains based on content type. Saved our team 15 hours a week.
That multi-modal stuff you want? Same approach. Build workflows that auto-process images, extract text, generate embeddings, and feed everything into your RAG system.
Your timeline’s realistic, but automation gets you there faster. Focus on systems that work without constant manual intervention.
Been working in GenAI for 18 months - your experience is pretty typical for most of us. Pure AI research jobs are rare. Most roles mix traditional dev work with AI integration, which sounds like exactly what you’re doing.
One thing others haven’t mentioned much: get comfortable with evaluation metrics beyond accuracy. You need to understand perplexity, BLEU scores, and custom evaluation frameworks when optimizing RAG systems. I spend about 30% of my time analyzing model outputs and tweaking retrieval strategies.
LangSmith monitoring is smart, but don’t ignore cost optimization. LLM calls get expensive fast in production. Learning to balance performance with token usage separates junior from senior AI engineers. I’ve seen projects tank because nobody tracked inference costs.
Your three-month timeline is doable but depends on your company’s AI maturity. If they’re still figuring out basic deployment pipelines, you’ll spend more time on DevOps than model work. The hybrid skillset you’re building is exactly what the market wants right now though.