Experience
4+ years
Projects
22 shipped
Tech Stack
35 technologies
Posts
(9 items)// Blog posts, articles, and write-ups
AI did not reduce my workload but It increased my throughput
I’m not doing less - I’m doing more: more experiments, more automation, more “small” ideas that actually make it to production. Lately I’ve been rotating between Codex 5.3 and Opus 4.6. Both are strong, but they optimize for different failure modes and once you treat them that way, picking the right one gets easy. Codex 5.3: depth + correctness Codex doesn’t rush. Give it a complex problem and it goes deep. One time, I asked it to persist AG Grid state (columns, ordering, filtering, sorts the full setup) against a specific version. It fetched archived docs, dug into node_modules, tracing the relevant APIs, validating assumptions, built a plan. One shot. Done. Explore → Reason heavily → Verify → Completeness over speed Codex fixes root causes, not symptoms. Minimum changes, maximum effect. Opus 4.6: Fast builder with strong UI instincts Opus is different. It follows instructions precisely - and that’s genuinely powerful. It’s fast, excellent at UI (seriously impressive), great at tool calls & best-in-class for building things from scratch And Claude Code’s harness makes it even better. 💡 Pro tip: With Opus, be explicit about intent, not just what. Tell it why you want something and watch it fly. My decision rule •🎯 High cost of mistake → Codex first •🎯 Speed to visible output → Opus first Two models. Very different strengths. Yes, Prompting matters. So does knowing which model won't start making things up when it gets tired. #AIEngineering #AINative #opus #codex #AgenticAI
The Workflow Shift That Changed How I Build Software
I've been coding with AI for a while now, but this workflow shift actually changed how I think about development. I used to open an IDE or a CLI tool and pull up a ticket. It works. But recently I switched to Codex (https://chatgpt.com/codex) a chat-first interface where the entire left panel is just my active projects. Here's what makes it different: I'm usually juggling 3-4 repos at once. I prompt in one project, and while the agent processes, I switch to another and keep working. No tmux sessions. No tracking which terminal belongs to which repo. No mental overhead of "wait, which window was that build running in?" ⚡ All projects visible in one sidebar ⚡ Queue multiple prompts across repos ⚡ Run, test, view diffs, and commit directly from the interface The speed caught me off guard. I'm getting most of my work done this way now, not because the models got better (though they did), but because the UX finally makes sense for parallel workflows. I think 2026 is the year these UI-first AI coding tools go mainstream. CLI tools had their moment and they're still great for certain workflows, but managing terminal history and multiple tabs on a laptop starts to feel brittle. The real shift isn't the AI. It's learning when to take control and when to let it run. Many of us have gotten decent at prompting. We know how much context to give. We know which parts need human judgment. That's when the workflow becomes effortless. One reality check: this workflow amplifies both productivity and mistakes. You need stronger discipline around reviewing diffs because you're shipping faster than your brain can context-switch. The mental model shift is worth it-just don't skip the review step. #AIEngineering #SoftwareEngineering #DevTooling #AIWorkflows #DeveloperExperience
I Rebuilt My Portfolio in 5 Hours. Three Years Ago It Took 5 Days
I rebuilt my portfolio website in under 5 hours. Three years ago, it took me 4-5 days. Here's what I did differently. The experiment: Set up three parallel git worktrees. Gave Claude Opus 4.6, GPT-5.3 Codex, and Gemini the same task - redesign my portfolio using Anthropic's frontend skill. Asked each model for 3 completely different iterations. Got 9 designs in about an hour. The workflow: Gemini struggled. The UI had too many issues to be useful. Codex and Opus? Both impressive. Codex nailed the visual design - one iteration looked so good I immediately knew it could work. But it wasn't functional yet. So I took that Codex design and ran it through both models again: ⚡ Codex: Stayed fast and clean. The code was flawless, structured, minimal but effective. It delivered the MVP beautifully. ⚡ Opus: Went deeper. Built working shortcuts, proper navigation, mobile-responsive details. The decision: After 3-4 hours with each, I chose Opus. Spent extra credits refining it. (Thanks Anthropic for $50 credit) The level of polish shocked me. What surprised me: I didn't plan this website. I iterated into it. We're at a point where development isn't about building from scratch - it's about steering AI toward what you're looking for. 🌟 Live at sushilkamble.com #AICoding #SoftwareEngineering #AIEngineering #DevTooling
I Refactored 50,000+ Lines of Code in 4 Hours - Here’s the System That Made It Possible
Refactored 50k+ lines of code in 4 hours using parallel AI agents. Here's what actually worked for me. The problem? Files hitting 4-5K lines. AI couldn't hold enough context to make clean changes. Copilot was literally falling back to CLI commands instead of proper refactoring. Token burn was insane. The approach? I didn't just throw AI at the problem. I orchestrated it. Step 1: Analyzed the codebase structure and identified natural break points Step 2: Had AI plan the work into 4 tasks (3 running simultaneously, 1 sequential) Step 3: Used Git worktrees to give each agent its own isolated workspace Step 4: Named each agent, gave it clear boundaries, told it to reorganize, not rewrite logic ⚡ Agent 1: Store refactoring ⚡ Agent 2: Component breakdown ⚡ Agent 3: Utilities consolidation ⚡ Agent 4: Final integration (sequential) The result? Zero files over 1,000 lines. Proper separation of concerns. Every component isolated. The codebase went from "AI struggles with context" to "AI knows exactly where to look." What surprised me: 💡 CLI commands > MCPs for this. AI already knows exactly what to run. Used standard Linux commands to navigate. AI nailed it without fancy integrations. Now I can use instruction files (CLAUDE.md, AGENTS.md etc) to guide AI through the architecture. It knows the structure before touching a single file. The takeaway? It's not always about how powerful AI is. It's about how well you orchestrate it to solve problems at scale. #AgenticAI #AIWorkflows #Productivity
AI Finally Cracked UI Testing - And It’s a Big Deal
UI testing was the one thing AI couldn't do good for me. But Now, I just wrote a prompt in plain english and it, just verified login, click through 7 buttons, and reported back in under 2 minutes. When you're iterating fast with AI, manually clicking through features becomes the bottleneck. Then I discovered agent-browser (https://agent-browser.dev/) by Vercel. It's a browser automation CLI built specifically for AI agents. What makes it different: ⚡ Uses element refs (drastically less context) ⚡ Zero configuration needed ⚡ Built in Rust for instant command parsing ⚡ Works with any AI agent through simple CLI commands I gave it a complete workflow (See in the video): - Login with credentials - Navigate to a specific workflow page - Click through multiple buttons - Run the workflow - Extract execution logs and final output Under 2 minutes. Fully automated. Zero manual intervention. What hit me wasn't the speed. It's that we finally have a verification loop. AI builds the feature. AI tests the feature. AI reports the results. Beyond testing, this unlocks end-to-end automation - from submitting weekly timesheets and pulling analytics across tools, to scraping job postings and bulk-updating product listings and many more thing - all driven by prompt that you can save. #agentbrowser #ai #automation #testing
Too Many Engineering Blogs. So I Built My Own Aggregator.
Lately, I’ve been getting into a habit of reading how teams at Uber, Netflix, Cloudflare, Microsoft, etc. design systems, scale infra, and solve hard engineering problems — mostly through their engineering blogs. But there was a problem: - Too many blogs - Too many tabs - Not enough time to read everything end-to-end What I really wanted was: “Give me the key ideas from all these posts in one place, and let me search across them.” So over a weekend, I built it. Enggist – Engineering, as a gist. ⚙️✨ 🔗 Link: https://lnkd.in/d7aeaDTF 💻 GitHub: https://lnkd.in/digECe8N - Fetches posts from a curated list of engineering RSS feeds - Stores them in a database - Runs a daily cron to summarize + categorize each post using AI (2.5 Flash, with guardrails) - Extracts key highlights + tags - Minimal frontend with fuzzy search Under the hood: - Cloudflare Workers for cron + ingestion + processing - Two cron jobs: one to fetch, one to summarize - Vercel for hosting and deploying 💰 Infra cost so far: ₹0 This didn’t start as a “learning project”. It started as: “I keep running into this problem. Let me build something that makes my own life easier.” The funny part is: once you commit to solving a real problem, the stack, the tools, the patterns… they quietly fall into place. Sometimes you don’t need a big plan. You just need a small pain, a free time, and the willingness to ship. 🚀😊
Django Rest Framework + React — Authentication workflow 2023 (Part 2)
This blog, we will go through the frontend part of the authentication workflow. We will create pages, components, custom hooks, utility functions, cookies handling, global user store & routes with route guard. By the end of this blog, you will have a full-blown, blazingly fast authentication workflow for your future full stack applications.
