From the Build Log
Technical deep-dives, architecture breakdowns, and real case studies from building AI systems in production. No fluff — just what actually works.
41 articles by Dhruv Tomar
Kimi K3 vs Claude Fable 5 vs GPT-5.6 Sol: Which One Should You Actually Pay For?
Kimi K3 is the largest open-weight model ever released, it leads the design arena outright, and it costs a third of what the leader charges. But it is 40% slower — and that single fact decides which model you should be paying for. Here is the comparison with the actual cost math, not just benchmark screenshots.
The Bodies Arrived First
In one week, a humanoid robot was decapitated in a combat ring and kept fighting, and another one held a glass without crushing it. The hardware problem is closing fast. The intelligence problem is not — and almost nobody reporting on robotics is telling you which one you're looking at.
Your Project on Localhost Doesn't Exist
A recruiter can't click localhost. Putting a frontend on the internet needs exactly two things — a domain and free hosting — and costs under ₹1,000 a year. The whole map, in three minutes.
The New Robot Hand Isn't the Story. The Word 'Teleoperated' Is.
1X gave its NEO humanoid a 25-joint, tendon-driven hand that can feel a glass start to slip. The spec sheet is real engineering. The viral demos don't say who's driving. Here's the read from someone who builds the software side.
Whisper Transcribed 60 Seconds of Silence as "Thank You."
I fed Whisper pure digital silence and it confidently returned "Thank you." Its no-speech probability said 0.000 — on silence AND on real speech. Here are the actual numbers I measured, and the two-floor gate that finally held.
Fable 5's 19-Day Ban Is Over. The Lesson for Builders Isn't.
Export controls froze the world's most capable public model for 19 days — then it came back with a classifier that blocks the exploit 99% of the time. What this teaches anyone building on frontier APIs.
Claude Fable 5 vs GPT-5.6 Sol: The Real Benchmark Story (It's Not One Winner)
Fable 5 leads the aggregate 91-86, Sol wins terminal work 88.8 vs 83.4, Fable owns SWE-Bench Pro at 80.3 — and costs double. The flagship choice is a workload decision, not a fan decision.
Open-Source Just Caught the Frontier: GLM-5.2, DeepSeek V4, Kimi K2.7 — What I'd Actually Run
GLM-5.2 trails Opus 4.8 by one point on hours-long coding tasks — MIT licensed, 1M context, free weights. The open-source tier list just got rewritten, and I run half of it daily.
The Mid-2026 Model Playbook: How I'd Set Up an AI Stack Today, From Scratch
Fable 5, Sol, Sonnet 5, Terra, GLM-5.2, DeepSeek V4 — the menu is overwhelming and the prices moved twice this month. Here's the exact decision framework I use to build production AI stacks.
Claude Sonnet 5 Just Flipped Agent Economics. Your Routing Table Is Out of Date.
Sonnet 5 scores within 3 points of Opus 4.8 on knowledge work at a fifth of the price — $2/$10 until Aug 31. If you run agents, your model routing just went stale.
GPT-5.6's Sol, Terra, and Luna: Model Naming Grew Up — and the Price War Got Real
OpenAI's three-tier Sol/Terra/Luna family went live July 9. Terra promises GPT-5.5 performance at half the cost. Here's the full mid-2026 pricing map — and how to route across vendors.
Claude Fable 5: Anthropic Just Shipped a Tier Above Opus. Here's What It Means for Builders.
Anthropic's new Mythos-class tier sits above Opus. Fable 5 is now the most capable Claude you can actually buy — but at $10/$50 per million tokens, when should you?
Claude Opus 4.8 Is the Coding Model Anthropic Should Have Led With
SWE-bench Pro up 5 points, terminal work up 8.4, and code that's 4x more honest about its own flaws. Opus 4.8's real upgrade isn't capability — it's trust.
MCP Goes Stateless on July 28. If You Run MCP Servers, Here's Your Migration Checklist.
The 2026-07-28 MCP spec removes protocol sessions entirely. Sticky routing, session stores, gateway inspection — gone. Here's what breaks and how to migrate.
30 Days With Claude Opus 4.7 in Production: 1M Context Wasn't the Story
I wired Claude Opus 4.7 into QuotaHit 30 days ago. The 1M context at standard pricing matters — but three quieter features changed my workflow more.
I Gave Claude Code Access to My Entire Business. Here's What Happened in 30 Days.
Not a sandbox experiment. I connected Claude Code to my CRM, codebase, email, Slack, deployment pipelines, and client projects — then let it run for 30 days. The results broke my assumptions about what a solo operator can do.
I Replaced 5 Hires With One AI System. Here's the Exact Stack.
A 20-person sales team needed 5 new hires. I built one AI system instead — pipeline scoring, lead routing, auto-follow-up, reporting. Here's exactly what I used.
My AI Setup Saves 15 Hours/Week Per Team Member — Here's How
15 hours per week. That's what every sales rep got back after I automated their grunt work. Not with one magic tool — with a system of 6 connected automations.
Stop Building Features. Ship Businesses.
The biggest mistake AI engineers make: they build features when clients need businesses. A feature is a login page. A business is a system that acquires, converts, and retains customers on autopilot.
The AI Stack I Use to Run 20+ Production Systems
Every tool, framework, and service I use across 20+ production AI systems — with why I chose each one and what I'd change today.
How to Build AI Agents That Actually Work in Production
Most AI agent tutorials end at 'hello world.' This guide covers the architecture, error handling, memory, and orchestration patterns that separate demos from production systems.
n8n vs Zapier vs Make.com: Why I Self-Host 218 Workflows
I've used all three. At 218 workflows, the cost difference is 50x. Here's an honest comparison with real numbers, real limitations, and when each tool actually makes sense.
RAG Architecture Guide: Building Retrieval Systems That Don't Hallucinate
RAG is simple in theory and brutal in practice. Chunking strategies, embedding models, retrieval tuning, and the reranking trick that cut our hallucination rate by 60%.
Claude Code Tutorial: How I Build 10x Faster with AI-Powered Development
Claude Code isn't just autocomplete. With skills, MCP servers, and custom agents, it becomes an AI development team. Here's my exact setup and workflow.
AI Automation for Small Business: What Actually Works in 2026
Forget the hype. Here are 7 AI automations that save real small businesses 15+ hours/week — with actual costs, setup time, and ROI numbers.
MCP Servers Explained: How AI Agents Connect to the Real World
MCP is the USB-C of AI — one protocol to connect any AI agent to any tool. Here's how it works, how to build one in 30 minutes, and why it changes everything.
How I Replaced 5 SaaS Subscriptions with One AI System
Jasper, Hootsuite, Calendly, Drift, and Clearbit — $847/month replaced with self-hosted AI workflows costing $35/month. Here's the exact setup.
Multimodal RAG: How to Build AI Search Across Text, Images, Audio, and Video
Text-only RAG is table stakes. Here's how I built a system that searches across PDFs, images, audio recordings, and video — all in one unified vector space.
AI for Indian Businesses: 7 Automations That Work Without Enterprise Budgets
Enterprise AI costs lakhs. These 7 automations cost under Rs.5,000/month and handle Hindi, WhatsApp, UPI, and Indian compliance. Built for real Indian businesses.
The Complete Guide to Deploying AI Applications to Production
From 'works on my machine' to 'running at scale.' Docker, AWS ECS, Vercel, environment management, CI/CD, monitoring — the complete deployment playbook for AI apps.
Building a Voice AI Agent: From Speech Recognition to Production Calls
Voice AI isn't a chatbot with a microphone. Latency budgets, interruption handling, accent robustness, and the telephony stack that handles 200+ real business calls daily.
How to Build a SaaS MVP in 2 Weeks with AI-Assisted Development
QuotaHit went from idea to deployed SaaS in 14 days. Here's the exact timeline, tech decisions, shortcuts that worked, and mistakes I'd avoid next time.
Prompt Engineering for Production: Beyond the Basics
Prompt engineering tutorials teach you temperature and few-shot examples. Production requires structured outputs, error recovery, cost optimization, and prompts that don't break when the model updates.
How I Built 7 AI Agents That Run an Entire Sales Department
QuotaHit isn't a chatbot — it's 7 autonomous agents handling lead gen, qualification, outreach, and closing. Here's the architecture that makes it work.
Voice AI in Production: Handling 200+ Calls/Day for a Construction SaaS
Angelina handles real customer calls for Onsite — a construction SaaS. Sub-1s latency, appointment booking, lead qualification. Here's what actually works in production.
39 Claude Code Skills in 30 Days: The Compound Learning System
From zero to 39 published skills on skills.sh. Not a tutorial — this is the system architecture behind building reusable AI capabilities at scale.
218 n8n Workflows: The Automation Architecture Behind Everything
218 workflows. 54 active in production. From social media posting to CRM hygiene to AI voice callbacks. Here's how I architect automation at scale.
Ghost Browser: Open-Source Browser Automation with Human-Like Behavior
LinkedIn auto-post, feed engagement, job applications, web scraping — all with human-like behavior that bypasses anti-bot detection. Architecture and anti-detection deep-dive.
From 0 to 10 Cr ARR: The AI Systems Behind Onsite's Growth
4 LangGraph agents serving 20 sales reps. Daily pipeline scoring of 500 leads. Facebook Ads optimization. The AI infrastructure that scaled a construction SaaS to 10 Cr ARR.
Building MCP Servers: How I Connected AI Agents to Everything
8 MCP servers for Euron, 2 for Onsite, custom servers for every client. MCP is how AI agents actually do things — here's the pattern I use to build them fast.
FitTrack AI: Running AI Entirely Client-Side for $0/Month Hosting
Camera-based rep counting, food photo analysis, workout tracking — all AI runs in the browser. No server costs, no API bills, no user data leaves the device.