Category: Programming | Reading time: 10 minutes
🇮🇩 Baca versi Bahasa Indonesia
In two earlier posts, I covered how to build a WhatsApp bot from scratch and how to connect it to an LLM to make it smarter. In this post, I want to step back and share the bigger picture: how those two techniques eventually became NEXI, the IT helpdesk bot that's now actually used in my day-to-day work.
This isn't a technical tutorial — it's more of a story about the process, the decisions, and the lessons learned from building something that has to be genuinely reliable, not just a demo.
The Initial Problem
At the manufacturing company where I work, the IT support team is small relative to 500+ employees. The questions coming in were repetitive: WiFi issues, password resets, printers not connecting, access requests. Most of them could be answered with nearly identical information — but each one still had to be answered manually, one by one, over WhatsApp.
I saw this as a clear opportunity for automation, and started with the simplest version first — exactly what I described in the basic WhatsApp bot article.
Evolving From Keyword Matching to an LLM
The first version of NEXI was just simple keyword matching. That was enough for the first few months, but it had a clear ceiling — every time a new type of question came in that hadn't been anticipated, I had to manually add a new rule. Eventually the keyword list grew into hundreds of lines and became hard to maintain.
The turning point was integrating Groq/LLaMA — the bot could suddenly handle a much wider range of questions without me constantly adding manual rules. But this wasn't a magic fix either; I still combine it with structured logic for anything that requires absolute accuracy.
The Challenge That Wasn't Obvious at First: Security
This is the part I think is most important but most commonly underestimated by people just starting a similar bot project. Once NEXI started being used for real (not just a demo), I realized there were plenty of security gaps that needed closing:
- Input validation — preventing users from sending commands or prompts designed to "trick" the bot into leaking information it shouldn't share with just anyone
- Rate limiting — preventing spam that could drain API quota or overload the system
- Access control — making sure the bot only shares sensitive information with numbers verified as employees, not with just any number that messages it
- Secure logging — recording activity for audit purposes, without carelessly storing sensitive data
I ran a thorough security review across several iterations, and managed to bring the number of critical vulnerabilities down from nine to zero. The process wasn't instant — each iteration usually uncovered new issues that weren't visible in the previous round.
The Biggest Lesson: A Bot That "Works" vs. a Bot That's Reliable
There's a big difference between a bot that works when you demo it to your manager, and a bot that's genuinely reliable when hundreds of employees rely on it every day. A few things I learned:
- Fallbacks are non-negotiable — if the LLM API goes down or is slow, the bot needs a backup path (a default response or escalation to a human), not just silence or an error
- Monitoring isn't optional — I need to know when the bot stops responding, not wait for a complaint from an employee
- Documentation matters for stakeholders — I created two separate presentation decks for management, explaining progress and the security results achieved, because non-technical stakeholders need a different lens than a commit history
The Outcome
NEXI is now part of the daily workflow at my workplace — reducing the repetitive load on the IT support team, with response times far faster than waiting for a human to become available. It's not a full replacement for the IT team, but it's an effective first filter for common questions.
Closing Thoughts
If you're just starting a similar project, my advice: start with the simplest version first (read the basics here), then move on to AI integration once the basic version is stable. And don't skip the security part — that's what separates a demo project from a system people can actually depend on.
If you need help building a similar system for your company or team — an IT helpdesk bot, automated customer service, or other internal automation — I'm open for freelance work. Check out the full NEXI and ARES case studies on my portfolio, or reach out via the Contact page.
Have questions about the process, or want to discuss your own use case? Leave a comment.
Tidak ada komentar:
Posting Komentar