Yash Khanvilkar (@Yash_KVR) header
Yash Khanvilkar (@Yash_KVR) profile picture

AI/ML | Robotics | Full Stack Dev

India
Joined February 2022
Photo by Yash Khanvilkar (@Yash_KVR)Photo by Yash Khanvilkar (@Yash_KVR)Photo by Yash Khanvilkar (@Yash_KVR)Photo by Yash Khanvilkar (@Yash_KVR)Photo by Yash Khanvilkar (@Yash_KVR)Photo by Yash Khanvilkar (@Yash_KVR)Photo by Yash Khanvilkar (@Yash_KVR)Photo by Yash Khanvilkar (@Yash_KVR)Photo by Yash Khanvilkar (@Yash_KVR)Photo by Yash Khanvilkar (@Yash_KVR)
Physical AI cheat sheet — sites and papers worth saving if you build robots that think. 1. Hugging Face LeRobot — datasets + policies for real robots 2. NVIDIA Isaac / Isaac Lab — sim → real training loop 3. ROS 2 + Nav2 — the boring backbone that still wins 4. Open X-Embodiment — cross-robot trajectory data 5. RT-2 / Octo / OpenVLA papers — VLA baselines to beat 6. Mujoco / Genesis — fast physics for iteration 7. Figure AI / Unitree docs — what hardware people actually buy Pattern: perception + policy + control + verification. Skip any one and the demo stays in the video. Save this for your next build weekend.
6
358
Your agent said password reset was done. Tests green. The reset link never expired and worked twice. An old email in an inbox was a key to the account. Fix: single-use tokens that expire in minutes, plus a test that reuses a link and expects a reject. Would you catch it?
23
Before you plug an MCP server into your coding agent, check these 7: Who maintains it A pinned version, not latest Every tool it exposes Read-only unless it needs writes A token scoped to one project No secrets in the config file A log of every tool call Bookmark this.
8
1
8
380
LLM tip: before you swap models in production, replay a sample of last week's real requests through both. Diff the outputs side by side, not just benchmark scores. Regressions show up in your traffic, not on a leaderboard. How do you test a model swap?
3
4
73
Your agent said single-use coupons were done. Tests green. It checked "used?", then marked it used in a separate step. Two checkouts at once both got the discount. Fix: one atomic update (WHERE used = false) plus a parallel redeem test. Do your tests run anything in parallel?
19
Your coding agent can leak your keys in 8 places most people never check. Traces, .env reads, MCP configs, commits, CI output, logs, command lines, pasted errors: each one with its fix, on one screen 👇 Bookmark this.
1
2
99
Coding agent tip for UI bugs: give it a browser. Chrome DevTools MCP lets Claude Code, Cursor or Copilot open the page, click, read console errors and take a screenshot. Now "fixed" means it saw the button work, not that the code looks right. What would you check first?
15
6
163
Your agent said file uploads were done. Tests green. It kept the filename the user sent. A name like ../../config could overwrite files outside the upload folder. Fix: name files on the server, and test a ../ upload expecting a reject. Do your tests send hostile filenames?
19
Before you let your coding agent run unattended, set these 7: A hard cap on turns or spend A timeout on every command A branch, never main Tests as the stop condition Network off unless it needs it A diff summary waiting at the end A ping when it stops or fails Bookmark this.
2
2
129
Claude Code vs Codex isn't either-or for a lot of devs anymore. Pick your split: A) Claude Code builds, Codex reviews B) Codex plans, Claude Code builds C) One agent for everything D) Something else (tell me) Which one is yours?
1
43
Your agent said the payment retry was done. Tests green. It retried the charge on timeout. The first charge had already gone through. Customers got billed twice. Fix: one idempotency key per order, so a retry returns the first charge instead of a new one. Would you catch that?
18
Before you trust a new coding agent harness, test these 7: 10 real tasks from your repo Pass rate, not vibes Cost per solved task Time per task Runs tests unprompted? Deny rules hold? Resumes after a crash? Same model, new harness, new results. Bookmark this.
1
1
3
119
Coding agent tip for bugs: don't ask for a fix first. Ask it to write a test that reproduces the bug and show it failing. Then fix it and show the same test passing. If it can't make the test fail, it doesn't understand the bug yet. Do you make your agent prove the bug first?
2
1
27
Your agent said the migration was done. Tests green. It renamed a column by dropping it and adding a new one. The test DB was empty, so nothing failed. In prod, every customer's phone number was gone. Fix: seed real rows, migrate, assert they survive. Does your CI test data?
1
31
Your agent says "done." These 8 tests check if it's true. Skipped tests, swallowed errors, secrets in logs, any user touching any order: each failure has one test that catches it. The whole sheet on one screen 👇 Bookmark this.
2
20
Claude Code tip: rules in your memory file are guidance. Hooks are enforced. Add a PreToolUse hook on Bash that exits with code 2 when a command matches rm -rf or push --force. The call is blocked and your stderr message goes back to Claude. What would you block first?
4
2
92
Before AI, even learning VS Code was part of the programming course 🤣🤣🤣🤣
5
3
16
440
Your agent said request logging was done. Tests green. It logged every request body. Passwords and auth tokens landed in plain-text logs. Fix: log an allowlist of fields, plus a test that sends a fake password and fails if it shows up in the logs. Would your tests catch that?
22
Before you run coding agents in parallel, set up these 7: A git worktree per agent Separate ports and test databases One task file per agent A token budget per agent Never two agents on one file One human who merges A review queue, not 6 chat tabs Bookmark this.
34
7 signs your coding agent's "done" isn't done: No test output in the transcript New tests marked skip Lockfile unchanged after a bump Errors logged, not raised Mocks where real calls should be Edits to files you never named "Should work" instead of "I ran it" Bookmark this.
2
1
3
200
Before a coding agent touches your repo, add these 7: One command that runs all tests Lint and typecheck commands A rules file it reads first A deny rule for .env and secrets Seed data, no prod creds A branch it can't merge itself A log of every command it ran Bookmark this.
3
7
235