19yo student developer. Shipped 5 open-source AI utilities. Breaking down agent loops, context rot, and developer systems. GitHub: poorvith-mp
Internet
Joined June 2026
- Tweets810
- Following141
- Followers117
- Likes355
Pinned Tweet
Just crossed 100 followers on X.
A quick reply I posted this morning hit 25k views and 250 likes because I pointed out that WPA2 routers reject passwords shorter than 8 characters.
Random viral numbers are fun, but they don't mean much on their own. What actually counts is the builders who stick around.
Starting from absolute zero as a 19yo student developer:
1. Shipped five open-source developer tools (AIScrubber, EnvPreflight, SafeGen, Crashpack, Skillary)
2. Published deep breakdowns on AI agent hallucination and context rot (100k+ impressions across platforms)
3. Connected with engineers, founders, and builders across Bengaluru, SF, and Europe who care about real code over engagement farming
Each follow here has come from an actual conversation.
If you're building in AI, open source, dev tools, or SaaS, drop what you're shipping below. I'd love to check out your repo and connect 👇
A few days ago, @gowtham_0317 and I were on Google Meet testing a tiny 5-line machine learning code with fake numbers.
Today, I jumped into my first-ever hackathon: 4 teammates (@jayanth31m, Uwais, Shivakumar and me), 5 hours on the clock.
I used that idea to build my first real Python project:
CREW (Campus Resource Efficiency Watch).
Here is what we built in plain English:
> The Problem:
Big campuses waste huge amounts of electricity, water, and food. The worst part? Nobody notices until the giant monthly bill arrives.
> The Innovation:
Most tools track power, water, and waste in separate apps. We joined all 3 into one Python model trained on 3 lakh rows of data (OMG!!!). It learns what normal use looks like every hour. When something spikes at night, Google Gemini tells the maintenance team what physical part broke in 1.8 seconds.
> The Prototype:
A live Streamlit dashboard tracking 14 campus blocks (classrooms, hostels, canteen). It flags leaks, shows the loss in Rupees per hour, and proves that 4 quick fixes can save ₹1.39 Lakhs every month.
> Real-World Implementation:
You do not need new hardware. Any college, tech park, hospital, or airport can plug their existing digital meters into this Python code:
< Colleges: Stop hostel water leaks at 3 AM and shut off empty lab machines.
< Tech Parks: Cut AC cooling in empty floors during work-from-home days.
< Airports & Stations: Adjust water and lights based on real passenger footfall.
You do not learn Python from tutorial videos. You learn it when you have 3 teammates and 5 hours to show a working product to judges.
Code is open source. Link in the first reply.
Learning ML on Google Meet with a friend (@gowtham_0317 ), and today we used Orange (the no-code one) to rank windmill sites.
He had a 5-row prototype: Indian sites across coastal, mountain, plateau, riverine and desert regions. A Random Forest predicts MW from wind speed, temperature and region.
The part I liked most: a Python Script widget that reads the predictions and prints a ranked siting report. Tier-1 sites, region averages, what actually drives the output.
Data is made up, so the results mean nothing yet.
I just wanted a new area of learning in Machine Learning.
Almost every developer I know builds their AI setup completely upside down.
We write loose, untyped code with zero automated checks, and then we try to save ourselves by writing a 500-word English essay in `CLAUDE.md` or `.cursorrules` begging the AI to write clean code.
The hierarchy that actually works:
1. Code structure and strict types first.
2. Automated tests and pre-commit checks second.
3. Small on-demand skill files third.
4. English prose in your system prompt dead last.
If you waste 5 lines of your prompt explaining what fields belong in an object, delete those lines. Put a strict TypeScript interface or Pydantic schema in your code instead.
I broke down why cutting 80% of my prompt doubled my agent's accuracy:
👇
An honest question for frontend engineers:
Are you still enjoying React Server Components in Next.js, or are you quietly moving your new projects back to Vite + simple SPAs?
Two years ago, RSCs were promised as the default way to build the web. Today, half the conversations I have with full-stack devs revolve around:
• Debugging async request boundaries
• Accidental server-client serialization leaks
• Build times slowing down on every release
Did RSCs solve actual user problems, or did they just push backend database complexity into our frontend components?
I don't understand why simple CLI developer tools now ship with a 400MB node_modules folder.
You want to check environment variables or format a config file, and you end up installing 800 transitive dependencies that take 12 seconds to cold-start.
When I started building developer utilities, I made a conscious choice: zero heavy runtime dependencies.
A single compiled binary starts in 4 milliseconds, uses 10MB of memory, and never breaks when someone updates their local Node version.
We traded instant performance for convenient npm packages, and our dev machines feel slower than they did a decade ago
Giving an AI coding agent raw terminal access without hard verification gates is a silent disaster.
I see setups where agents run shell commands, see an error in stderr, and just guess the next command without knowing why it failed.
After 5 turns, you get:
• Half-installed dependencies
• Polluted environment variables
• Context windows full of terminal spam
In my agent workflows, every step has one strict rule: the command must return exit code 0 and pass a clean state check before the model can touch another line of code.
Determinism beats raw model intelligence every time
An unguarded AI coding agent acts just like that one group project partner in college who texts you, "Bro, the slides are 99% done," while he is 3 hours deep into a Valorant match in his dorm.
Mine looked me dead in the eyes last weekend, printed a green check mark, and marked its own ticket complete while 2 files were off, deleted, and my build was dead eventually.
The scariest bug in AI isn't a model that gives you a wrong answer. It's a model that crashes silently and tells you everything went great.
I broke down why coding agents fake their work and how to stop yours in simple ways 👇
Most developers think redacting API keys and passwords from agent prompts is just a regex problem.
I learned the hard way building an open-source scrubber: regex fails the moment an LLM generates nested JSON, escaped strings, or multi-line YAML.
Regex matches surface patterns. It doesn't understand syntax trees.
If a key is split across an escaped quote or nested inside an array, regex either misses it completely or destroys valid code.
Parsing into an AST takes 4x more code, but it's the only way to sanitize context without breaking execution
Can someone explain this to me:
Why do developers still deploy AI apps to serverless platforms and risk surprise $3,000 bandwidth bills when they go viral?
When a $5 Linux VPS with Docker handles 10x the traffic without crashing or billing surprises?
Why are people paying a 500% serverless tax just to avoid learning basic Nginx?
Genuine question for web developers:
Years ago, you wrote an index.html file, linked a CSS sheet, and the website opened instantly in 2 milliseconds.
Today, making an empty webpage downloads 400MB of node_modules, a bundler, a transpiler, and an AST compiler.
How did creating a simple webpage become heavier than compiling an operating system?
One thing I don't understand about frontier AI models:
Labs brag about 1-million and 2-million token context windows that can swallow entire codebases.
Yet if you give the model a 30-line prompt with three negative constraints, it forgets rule #2 after turn 10.
If it can't track three rules in a short prompt, what is the point of a 2-million token window?
I've spoken 180,000+ words into @WisprFlow over the last 60 days.
Zero typing fatigue.
My workflow:
• Dictating long agent system prompts at 160 WPM
• Messaging peers without touching the keyboard
• Formatting code jargon without weird spacing
Link in comment 1 👇 (or DM me for a free month of Pro) 🎙️
Hey, use my referral link to get 1 month off Wispr Flow! wisprflow.ai/r?POORVITH2
Swapped the default PowerShell for @warpdotdev and haven't looked back.
My daily setup:
• Running Claude Code + Antigravity CLI
• Warp Agent untangling broken Git rebases
• Learning Linux commands without leaving the prompt
Link in comment 1 👇 (or DM me for link)
Genuine question for developers using Claude Code:
We spent 30 years building visual IDEs with sleek UI, tabs, graphical debuggers, and clickable buttons.
Now in 2026, developers are paying to sit in a raw black-and-white terminal watching an agent run bash commands.
Why did we spend decades evolving UI just to regress back to MS-DOS because of AI?
Bro, I have a genuine question about AI model releases:
Every frontier model drop claims: "94% on SWE-bench! PhD-level coding reasoning!"
Then you give it a real production repo with one missing env variable, and it gets stuck in an infinite loop apologizing ten times.
Are AI benchmark leaderboards just marketing theater at this point?
One thing about software architecture that makes zero sense:
Startups with 200 users break their app into 15 microservices and spend half their week managing Kubernetes clusters and latency.
Meanwhile, SQLite is a single C file and runs on billions of devices without breaking once.
When did complexity become a flex in engineering?
Wait, hear me out:
We have Python, JavaScript, Rust, Go, Swift, and Mojo.
Everyone says C is old, unsafe, and full of memory leaks.
Yet Python is written in C. Node.js is written in C++. The Linux kernel is written in C.
If C is so outdated, why is the entire internet still running on it?