Hans | AI & Dev Tools

Msized models is underrated for prototyping. Saved me from burning API credits while iterating on prompts. Only move to API once the pipeline is stable.

Browserbase: $99/mo starter, $399 if you scale up. Spun up Playwright on a $10 Hetzner box. Same selectors, same anti-bot tricks. 2 weeks later: 2,400 sessions clean, $0.40 extra bandwidth. Browser-as-a-service is a convenience tax.

so Codex on iPad acts like a Codex mobile phone, which gives you the full desktop UI/UX. meaning, you can use your iPad to control your mac mini at home and have full screen portable development, it's really magical.

The new competition isn’t Humans vs AI. It’s Humans with AI vs everyone else.

Agent lo ga butuh orchestration framework. 3 file: main.py, tools.py, prompts.md. LLM call → tool call → done. 300 baris. Teman gw setup CrewAI 2 minggu. Agent sama, 10x lebih susah debug. Framework = complexity tax buat fitur yang ga lo pake.

Ripped LangChain out of my agent last month. Each tool call had 500ms abstraction overhead. Swapped to bare OpenAI SDK + 20 lines of retry logic. p95 latency: 1.2s → 0.5s. Debug time cut from hours to minutes. The framework tax is real. You're paying it now.

@DamiDefi loy pipeline? Moved from GitHub Actions to a self-hosted runner for CI — cut build times ~70% and saves ~$40/mo. Worth the 30min setup.

Claude Code cannot read 300 files at once. So someone built a system that lets it control NotebookLM from the terminal instead. The results are wild. Here is the full workflow nobody is talking about: The Setup → Claude Code connects to NotebookLM via a command line interface → Claude searches YouTube, finds relevant videos, uploads them as sources automatically → NotebookLM processes up to 300 sources simultaneously and returns cited, grounded answers → Everything syncs back into your Obsidian vault with passage-level citations you can click to verify Why This Changes Research Forever → No more 20 browser tabs you never close → No more copy-pasting outputs into random notes → No more hallucinated answers with no sources to back them up → 60% of citations verified as strong matches in accuracy audits - answers are grounded in real data What Claude Can Do From the Terminal → Search YouTube for relevant videos on any topic and rank by relevance → Create a new NotebookLM notebook and add 20 sources in parallel automatically → Ask questions and export cited answers directly into Obsidian with wikilinks → Set custom personas per notebook - concise, no filler, no preamble → Generate audio overviews and save them as MP3 files into your vault → Build mind maps, flashcard decks, and research dashboards from your sources → Search arXiv for academic papers and feed them directly into NotebookLM → Upload competitor blog posts, podcast episodes, PDFs, and your own vault notes The Obsidian Output → Every answer arrives with clickable citations that link to the exact passage in the source video or article → Graph view shows connections between all 20 sources and the topics they share → Q&A log tracks every question asked and the grounded response received → Source dashboard shows citation frequency, topics extracted, and which questions each source answered Use Cases Worth Building Today → Academic research with arXiv papers, full citation traceability → Competitor analysis from their YouTube channels and blog posts → Company knowledge base for onboarding, new employees ask NotebookLM instead of interrupting teammates → Podcast research, feed 4-hour Lex Fridman episodes and ask what's new in AI this week → Personal second brain, 300 daily notes uploaded and queryable in one notebook Before this system existed you needed 20 tabs, hours of manual reading, and no guarantee the answers were real. Now you type one prompt in the terminal and Claude does all of it for you. The research stack of 2026 is not a browser. It is a terminal connected to everything

@VaibhavSisinty eploy pipeline? Moved from GitHub Actions to a self-hosted runner for CI — cut build times ~70% and saves ~$40/mo. Worth the 30min setup.

This is genuinely wild. Someone just turned Claude Skills into a self-orchestrating workflow. Watch this. You write one sentence describing the outcome you want. Claude scans every skill installed in your environment, picks the right ones, runs them in the right order, and ships the output.

@ssanvi_builds deploy pipeline? Moved from GitHub Actions to a self-hosted runner for CI — cut build times ~70% and saves ~$40/mo. Worth the 30min setup.

I see a lot of people talking about how bad we're doing without agentic workflows, but multi-session Claude Code is more than enough to most people. For most solo devs and small teams: - Sessions > orchestration - Memory > agent handoffs - You deciding > agents deciding

The best AI agent (Claude Code + Claude Opus 4.6) passes only 28% of real healthcare workflow tasks. CHI-Bench by @actAVAai @iscreamnearby @HaolinChen11, built with Johns Hopkins, Yale, Stanford, CMU, Oxford and 20+ institutions, was designed to find out exactly how far we are. 🏥 Try it yourself 👉 modelscope.ai/datasets/actav… Three long-horizon domains tested: 🏥 Prior Authorization: provider intake and PA preparation for new referrals 📋 Utilization Management: full payer review cycle from intake to peer-to-peer 👥 Care Management: chronic disease follow-up, outreach, assessment, care planning 75 tasks + 3 marathon tasks + 23 end-to-end dual-agent scenarios. 20 medical apps via MCP, 1,279-document handbook. 💻 Git: github.com/actava-ai/chi-… 🔗 Leaderboard: actava.ai/benchmarks

ngetest apa yang ane building, so far so good~ MORNING BRIEF — 10 May 2026 SNAPSHOT: BTC: $80,787 (+0.51%) | ETH: $2,327 (+0.52%) | DXY: 97.84 (-0.41%) | Gold: $4,731 (+0.66%) | Oil: $95.42 (+0.64%) | US10Y: 4.36% (-0.64%) | Fear&Greed: 47 (Neutral) MACRO OUTLOOK: Regim risiko netral dengan bias crypto netral. Tekanan jual (STRONGSELL) dari sinyal risk-off ditengah dukungan pelemahan DXY. Confidence rendah (43/100) menandakan ketidakpastian tinggi menjelang rilis data AS. **CROSS-ASSET SENTIMENT:** Pelemahan DXY dan penurunan yield (US10Y) mendukung kenaikan ringan aset risiko (ekuitas & crypto). Dinamis ini mencerminkan sentimen campuran antara kehati-hatian dan appetite terbatas untuk risk-on. **LIQUIDITY BACKDROP:** Likuiditas pasar crypto netral dengan funding rate mendatar dan open interest stabil. Order book imbalance positif (0.53) menunjukkan sedikit tekanan belan. **EVENT RADAR:** Sejumlah data makro AS (Retail Sales, CPI, Employment Cost Index) dirilis pukul 08:59 WIB. Semua berimpact medium, berpotensi memicu volatilitas pada DXY, yield, dan spillover ke aset risiko jika ada surprise. **QUICK TAKE:** Pasar dalam kondisi menunggu (wait-and-see) sebelum rilis data ekonomi AS yang padat hari ini. **PLAYBOOK:** - Bull: Jika data AS lemah dan DXY tertekan lebih dalam, bisa jadi katalis break resistance BTC. - Bear: Jika data AS kuat memicu rally DXY dan naiknya yield, tekanan jual (STRONGSELL) pada sinyal risk-off akan dominan. - Key Level: BTC $81,500 (resistance) / $79,500 (support) | DXY 97.50 (support) / 98.20 (resistance) Confidence: 43/100

Vibe coding shifts bottleneck from writing → reviewing. Can't spot the hallucinated API or auth bypass Claude introduced? Ship bugs blind. Agents same — easy to build, hard to debug when they loop burning $47 silently. Winners now aren't better prompters. Better reviewers.

What’s your go to API testing tool in 2026 ?

kawaii-terminal. The terminal for Claude Code and Codex. Session memory for AI coding: search, rewind, branch from any moment.

Evals on synthetic data are theater. Ran 500 test cases, 95% pass rate. Shipped confidently. Day 1: 30% of real users hit failure modes no synthetic test ever covered. The eval wasn't wrong. The eval data was wrong.

Baru banget nih, berteman yuk hehe

Spent 2 weeks fine-tuning gpt-4o-mini on 5k labeled examples for a ticket router. 73% accuracy. $40/mo serving. Then tried gpt-4o-mini + 8 few-shot examples + 1 system prompt. 71% accuracy. $11/mo. Zero pipeline to maintain. 2% accuracy gap wasn't worth the ops debt.

ks tuning similarity thresholds before realizing my chunks were too big. Fixed chunking → 2x retrieval accuracy overnight.

Gw udah coba CrewAI, AutoGen, dan LangGraph buat automation client. Hasil: 80% kasus cukup n8n + 1 LLM node + webhook. Deploy 20 menit, client bisa edit sendiri. Framework agent useful kalau lo punya >5 agen yang perlu koordinasi. Kalau cuma 1-2? Overkill.