Vikash Singh Tomar retweeted
Everyone’s building AI features.
Almost nobody is testing them.
With normal code, a test passes or fails.
With AI:
-> Same input, different output.
-> “Works” depends on who’s reading it.
-> It’s fine in the demo, wrong in production.
If you can’t measure whether your AI feature got better or worse after a change, you’re not engineering it.
You’re hoping.
Vikash Singh Tomar retweeted
SELECT COUNT(*) FROM users;
SELECT COUNT(email) FROM users;
Same table. Same rows.
Two different numbers.
No bug. No error.
Why? 🤔
Vikash Singh Tomar retweeted
(1/3) - AIBEAT
An open-source framework for testing LLM security and agent behavior, with PromptBeat for adversarial testing and AgentBeat for tracing what actually happens during execution.
Vikash Singh Tomar retweeted
Backend developers don’t get paid for writing code.
They get paid for what happens when the code meets reality.
Reality is:
→ Users who click twice.
→ Networks that drop mid-request.
→ Data nobody told you could be empty.
→ Traffic nobody planned for.
Anyone can write the happy path.
The job is everything else.
Vikash Singh Tomar retweeted
Same paste. Two outcomes.
Left: your API key goes straight to an AI tool.
Right: SecureIntent catches it before you hit send.
On-device detection. Chrome & Firefox. Free to start.
Which developer are you? 👇
secureintent.ai
Vikash Singh Tomar retweeted
Me: let's build an AI agent to save 10 hours a week
The agent: reads .env, prints my API key, pushes it to logs
Me: spends 12 hours rotating credentials and investigating the incident
AI is incredible 💀
Vikash Singh Tomar retweeted
Asked at Microsoft, Databricks, Glean & Palantir 🤔
You’re deploying a 70B-parameter LLM in production.
How much GPU memory do you actually need to serve it?
Vikash Singh Tomar retweeted
Your Tech Lead says:
We don't need Redis. The database can handle the traffic.
What's your response? 👀
Vikash Singh Tomar retweeted
Interviewer:
If an LLM's training data ends in 2024, how can it answer questions about events in 2026?
Vikash Singh Tomar retweeted
Senior Developer says :
DELETE /users/{id}
return 204 instead of 200 after successful deletion. You are a junior engineer
What's your answer ?
Vikash Singh Tomar retweeted
Interviewer:
If I fine-tune an LLM on my company's data, do I still need RAG?
Vikash Singh Tomar retweeted
Junior dev: "I'll just quickly fix this bug."
3 hours later
git commit -m "fix"
git commit -m "fix again"
git commit -m "actually fixed"
git commit -m "ok NOW it's fixed"
git commit -m "please work"
git commit -m "I give up"
git commit -m "found it, it was a typo"
The bug: usre.name instead of user.name.
Seven commits. One missing letter. 😭
Vikash Singh Tomar retweeted
Your environment variables are encrypted.
Your deployment platform gets compromised.
Some secrets may still be exposed.
Railway's latest security incident is a reminder:
Encryption matters. But credential security doesn't end there.
If you use Railway, rotate your secrets and review your API usage.
Don't wait for an unexpected bill to find out your key was exposed.
Official source: blog.railway.com/p/security-…
Vikash Singh Tomar retweeted
Interviewer:
If I increase an LLM's context window from 32K to 1M tokens, does that mean it can reliably remember everything in that context?
Vikash Singh Tomar retweeted
You build a food ordering app.
Friday 8 PM: dinner rush.
100,000 users open the app at the same time.
Every user hits the "Restaurants near me" API.
Same query. Same area. Same results.
100,000 times.
Database is sweating.
Restaurants list changes once every few hours.
How do you handle this load without upgrading the database? 🤔
Vikash Singh Tomar retweeted
Database Interview Question:
Your database replication lag suddenly jumps to 30 seconds, but the primary database is healthy.
What could cause this, and how would you troubleshoot it without taking the system offline?
Vikash Singh Tomar retweeted
Backend developers:
For GET /api/products, which response format would you ship to production?
Why?
Vikash Singh Tomar retweeted
Your AI agent has access to your terminal, your files, your API keys and your production logs.
but sure, let's worry about whether ChatGPT is stealing our jobs😂
Vikash Singh Tomar retweeted
Genuine question:
If your application uses JWT authentication, how do you invalidate a token immediately after logout without maintaining a server-side blacklist or session store?
Is truly stateless authentication really practical in production?






