Quartz 4

thoughts on ai engineering

6 min read

2/16 - 2/17/26

thoughts (+ my place)

  • the reason i did a bootcamp!
    • šŸ£Spring & SpringBoot, Backend Engineering
    • back then, i thought that AI related fields only accepted people with at least master degrees, so i wanted to switch to a more traditional ā€œsweā€ roleā€¦ and i didnā€™t have much depth in software engineering so i chose to do a backend (springboot) bootcamp.
      • I could have done a full stack or a frontend course, but for me personally i was more intrigued by the problems backend had to face - scalability, maintaining servers & requests, cloud/infra stuff, writing apis etc
    • But now, I think AI and backend engineering synergizes well together, and maybe I donā€™t have to pick one of them but mesh them into 1 and do both? so right now i think i am in a good spot to position myself as an AI-driven backend engineer (whatever this implies lol)
  • what exactly is ai engineering? how does it fit so well with backend engineering?
  • an interesting point 1
    • a redditor from this comment mentioned that nobody was an ā€œai engineerā€ literally few years ago - theres so many jargon and AI terminology thrown around that normal SWE had difficulty understanding them. But as soon as you get into it and really understand the concepts, LangGraph or Pydantic Ai or other AI frameworks becomes rlly easy + straightforward
        • understanding this is what makes you replaceable by someone who understands this and uses AI
        • they also build all AI logic with langgraph (they think there is no more convenient way to build sophisticated AI backend as it handles everything - monitoring, deployment etc)
  • an interesting point 2
    • i talked with my friend jenny (rn swe in microsoft), and i asked about her current job/task/projects, and she mentioned that AI can really replace a lot of the manual labor of writing code, but it still needs a human-in-the-loop to guide it because it doesnā€™t have good judgement deciding things, but it does seem inevitable that SWE will become managing a ton of subagents
    • which also led me to wonder about the future of SWE because, companies are already not hiring juniors but instead, they are just adding more ā€œjunior AIsā€ that the seniors can control to do the work of human junior devsā€¦ (what is going on in the cs industry?)
    • as for my career, traditional swe will be really hard to break in esp for juniors/new grads, so im planning to grind both ai engineering + backend so that my resume will be more noticed :ā€˜))

which tools?

  • LangChain
  • LangGraph + Pydantic AI
  • some inspirations
    • this comment from reddit: this dude specialized in a LOT of AI stuff AND software/enterprise stuff. Their advice is to learn programming patterns and learn to write good and clean code. ā†’ singletons, inheritance, factory pattern, etc.. they reappear everywhere. And they realized that langchain was actually useless and overcomplicated
      • reference recs: ā€œuncle bobā€ (j google that)
  • some tips? (from a reddit comment)
    • agent specificity - donā€™t make generic agents. Donā€™t make 1 agent to rule them all - thatā€™s a fools errand (also the #1 mistake from ppl!!)
    • all agents need exit ramps - give them the ability and direction to kick something to a human if they can complete a task
    • robust resumption rules - agents can and will get disrupted by unexpected situations. The ability to stop, resume by getting back up to speed is critical to success
    • plan, plan, plan - a shitty directionless plan will result is subpar results.
    • others
      • Frameworks that lock in control flow often miss decision points, get stuck in loops, go down rabbit holes that canā€™t continue.
      • Frameworks that donā€™t regulate flow, but focus on robust conversation management/handoff management often terminate prematurely, donā€™t complete, or meander inefficiently.
      • Master single agents first, do your own manual handoff between agents so you can control flow and data handoffs, so that you understand what is likely to happen, before trying a multi-agent ecosystem.
      • Once an agent tilts the system to an undesirable direction gets magnified with each subsequent agent interaction. Itā€™s the old game of telephone that you need to combat.

domain knowledge?

A combination of AI and backend engineering

  • AI, ML, DL fundamental knowledge
  • ai engineering
    • read the AI engineering book (chip huyen)
  • Backend engineering as a field
    • ā˜•Java for backend
    • the important and basic concepts
    • database
    • distributed systems
      • designing data-intensive applications (book)

Graph View

Backlinks