I tested 5 vector databases using 1M embeddings + real queries.
Results surprised me
Fastest: Pinecone & Qdrant
Most scalable: Milvus (GPU)
Best filtering: Qdrant
Best hybrid search: Weaviate
Best for prototyping: Chroma
Check out my medium blog: @officialpreksha2166
🎥 Watch the demo: https://t.co/Fg0OTSzq4W
🌐 Try it: https://t.co/cxHN1ZCM52
💻 Code: https://t.co/RHI1fmKPQ2
Built with Chirag Patankar at ForgeHacks 2026.
#MachineLearning#Cybersecurity
Introducing VisaShield — built at ForgeHacks 2026!
An AI + Cybersecurity platform that helps international students detect scam calls, phishing emails, fake ICE/USCIS threats & fraudulent OPT job offers.
Here's what's under the hood 🧵👇
https://t.co/p5q23t51uc
#AI#Cybersecurity
📊 6/ Evaluation matters!
On our real-world test set of 56 messages:
• F1 score: 0.899
• Precision: 0.969
• Recall: 0.838
• False-positive rate: 5.3%
These are prototype results on a small test set, not a guarantee of real-world accuracy.
🧩 I built an Image Mosaic Generator that recreates photos using tiny cartoon animals, LEGO bricks & more! 🎨
Built with Python, NumPy, KD-trees & CIELAB color matching.
🎥 Watch the demo!
💻 GitHub: https://t.co/l73cu98Cco
#Python#ComputerVision
Got this back from the client today: "She brought our legacy business into the digital age smoothly and professionally." — Anil Jadhav, Smart Art proprietor. First real client, first real testimonial. Feels different than a GitHub star.
Handed it off to the client, who'll publish it on her own domain. Recording a full walkthrough next. First real client project, first real lesson: understand the customer before you write a line of code.
This is the first project I've built for a real external client instead of myself. The hardest part was never the React, it was figuring out what a 30-year offline business actually needed versus what a checklist says a website should have.
Content layer: 8 SEO-researched blog articles (Warli art, industrial floor marking, epoxy coatings), a 59-image gallery across 7 categories, full JSON-LD structured data and per-page meta tags generated programmatically.
Contact form: dual-channel submission (Google Sheets via Apps Script + WhatsApp), phone validation regex for Indian mobile numbers, live character counter. No backend server to maintain, and the owner can open a spreadsheet himself.
Made it a full PWA: service worker, offline caching, installable manifest, push notification support. Most of this business's customers are on mobile data, not desktop broadband, so this isn't a nice-to-have.
Just wrapped building a full production website for a real client: Smart Art, a 30+ year industrial art & signage business in Thane run by artist Anil Jadhav. Siemens has been a client since 1990. No spec doc, no PM. Thread on what I built 🧵
Stack: React 18, React Router 6, Framer Motion for transitions, React Helmet Async for per-page SEO, React Markdown for blog rendering. 9 pages, 32 components, lazy-loaded with code splitting to stay fast on mobile data.
The real constraint wasn't technical, it was that customers order over WhatsApp, not forms. So the site routes straight to WhatsApp for ordering instead of a contact form nobody would use. Build for how people actually work, not what looks impressive.