Patrick Phat N.

After months of learning by doing, I finally developed my own AI product builder workflow. It makes it straightforward to spot small daily friction, build focused tools to solve them, and eat my own dog food. I spent two days building FirstCustomer. It helps you find real people who already have specific pain points and are actively waiting for solutions. Instead of manual guesswork, it scans community signals to connect you with high-intent early users. The web app is live, and the mobile version will be out soon. https://firstcustomer.net/ #BuildInPublic #AIAgent #AIProductBuilder

7

🚀 𝐏𝐚𝐩𝐞𝐫 𝐀𝐜𝐜𝐞𝐩𝐭𝐞𝐝 𝐚𝐭 𝐄𝐀𝐂𝐋 2026 (𝐅𝐢𝐧𝐝𝐢𝐧𝐠𝐬)! I’m excited to share that our paper has been accepted at 𝐅𝐢𝐧𝐝𝐢𝐧𝐠𝐬 𝐨𝐟 𝐀𝐂𝐋: 𝐄𝐀𝐂𝐋 2026 📄✨ 🔗 Paper Link: https://lnkd.in/gkUTuU93 🎯 𝐏𝐫𝐨𝐛𝐥𝐞𝐦 With the rapid growth of academic video content, accessing relevant information efficiently has become challenging—especially for 𝐁𝐚𝐧𝐠𝐥𝐚, a low-resource language. Existing solutions: • Focus mainly on 𝐄𝐧𝐠𝐥𝐢𝐬𝐡 • Work on 𝐅𝐨𝐫𝐦𝐚𝐥 𝐭𝐞𝐱𝐭, not 𝐬𝐩𝐨𝐤𝐞𝐧 𝐚𝐜𝐚𝐝𝐞𝐦𝐢𝐜 𝐜𝐨𝐧𝐭𝐞𝐧𝐭 • Lack 𝐞𝐧𝐝-𝐭𝐨-𝐞𝐧𝐝 𝐯𝐢𝐝𝐞𝐨 𝐬𝐮𝐦𝐦𝐚𝐫𝐢𝐳𝐚𝐭𝐢𝐨𝐧 𝐩𝐢𝐩𝐞𝐥𝐢𝐧𝐞𝐬 • Struggle with 𝐀𝐒𝐑 𝐢𝐧𝐚𝐜𝐜𝐮𝐫𝐚𝐜𝐢𝐞𝐬 and contextual coherence 💡 𝐎𝐮𝐫 𝐒𝐨𝐥𝐮𝐭𝐢𝐨𝐧 We propose the 𝐟𝐢𝐫𝐬𝐭 𝐞𝐧𝐝-𝐭𝐨-𝐞𝐧𝐝 𝐩𝐢𝐩𝐞𝐥𝐢𝐧𝐞 for 𝐁𝐚𝐧𝐠𝐥𝐚 𝐚𝐜𝐚𝐝𝐞𝐦𝐢𝐜 𝐯𝐢𝐝𝐞𝐨 𝐬𝐮𝐦𝐦𝐚𝐫𝐢𝐳𝐚𝐭𝐢𝐨𝐧: 🔹 Audio preprocessing to improve transcription quality 🔹 Speech-to-text using optimized ASR 🔹 Smart 𝐜𝐨𝐧𝐭𝐞𝐱𝐭-𝐚𝐰𝐚𝐫𝐞 𝐜𝐡𝐮𝐧𝐤𝐢𝐧𝐠 for long transcripts 🔹 Fine-tuned 𝐁𝐚𝐧𝐠𝐥𝐚𝐓5 for abstractive summarization 🔹 Fine-tuned 𝐦𝐓5-𝐦𝐮𝐥𝐭𝐢𝐥𝐢𝐧𝐠𝐮𝐥𝐚𝐥-𝐗𝐋𝐒𝐮𝐦 for title generation 🔹 Timestamp-integrated summaries for efficient navigation 📊 𝐊𝐞𝐲 𝐂𝐨𝐧𝐭𝐫𝐢𝐛𝐮𝐭𝐢𝐨𝐧𝐬 ✅ New benchmark datasets (10𝐊+ 𝐬𝐮𝐦𝐦𝐚𝐫𝐲 𝐩𝐚𝐢𝐫𝐬, 1𝐊+ 𝐭𝐢𝐭𝐥𝐞 𝐩𝐚𝐢𝐫𝐬) ✅ Strong performance outperforming existing models ✅ Improved accessibility of educational content in 𝐁𝐚𝐧𝐠𝐥𝐚 ✅ Robust pipeline of video summarization with cross-domain generalization 𝐆𝐫𝐚𝐭𝐞𝐟𝐮𝐥 𝐭𝐨 𝐦𝐲 𝐚𝐦𝐚𝐳𝐢𝐧𝐠 𝐜𝐨-𝐚𝐮𝐭𝐡𝐨𝐫𝐬 Farhatun Shama, Abdul Aziz, Md Kaykobad Reza, Md Shahidul Salim (Shakib) for their collaboration, support, and contributions throughout this work. This work aims to bridge the gap in 𝐥𝐨𝐰-𝐫𝐞𝐬𝐨𝐮𝐫𝐜𝐞 𝐍𝐋𝐏 and make academic content more accessible for Bangla speakers. #NLP #BanglaNLP #TextSummarization #EACL2026 #Research #MachineLearning #AI

53

3 Comments

As a Tech Director, I see Product Managers - PMs getting pulled in ten different directions every day. Operations wants this, Marketing needs that, and the CEO just saw a new AI feature on X. The natural instinct is to say "Yes, we’ll look into it." We want to be helpful. We want to be "agile." But here's the reality: A PM who says yes to everyone is a PM who builds a fragile, unsustainable system. In our weekly sync, we discussed the "Ruthless Prioritization" mindset. It’s not just about picking what to do; it’s about having the courage to communicate delays and say "No" to things that don't move the needle. Sustainable engineering is about working on fewer, more impactful things and having the backbone to protect the team's focus. Your job isn't to manage a backlog. It's to protect the vision.

22

Agent can really get lost in the product. Recently, Duc Ngo and I started a hobby to evaluate agent interfaces. We believe that agents will be the main user of any products, and agent experience designer will be an impactful role. Agent interaction should be understood from the product side. Even a stronger model can get confused when a product is badly designed. For example, when we tested one case with Metabase MCP V1, the agent searched for and found a metric. However, the result did not give it a usable address for that metric. It is like finding a restaurant on Google Maps but not getting its address. Any human would be confused too 🤣 Metabase MCP V2 has fixed this and is much friendlier for agents to use. That is a small issue, but definitely a bigger lesson: a good interface is not only about returning the right answer, but also guiding the agent toward the next clear steps. We are evaluating products everyday, and will share more learnings frequently.

28

2 Comments

The solo startup stack in 2026. ✦ n8n — automate everything, self-hosted, $0 ✦ Supabase — backend, auth, database, storage, edge functions ✦ Upstash Redis — caching and rate limits ✦ Cursor — AI-assisted coding ✦ v0 — scaffold UI in seconds ✦ Bolt — working prototype in hours ✦ Claude — architecture, PRDs, debugging logic ✦ Vercel — deploy frontend and serverless instantly ✦ Railway — containers when Vercel isn't enough ✦ Stripe — payments, always ✦ Lemon Squeezy — Stripe but with tax handled for you ✦ Resend — transactional emails ✦ Loops — marketing email sequences ✦ Framer — landing page that actually looks good ✦ PostHog — product analytics + session replay ✦ Microsoft Clarity — free heatmaps ✦ Cloudflare — security, CDN, DDoS protection ✦ shadcn/ui — components that don't look like garbage ✦ Tailwind CSS — styling; nothing else comes close I hope this helps you. Follow me, Ankit Suthar for more daily posts on AI productivity tools, tutorials, and updates

4