Dreamer with a head full of ideas. Kumar Manik - The next big thing.

Day 27 - 180 Days of becoming a better AI Engineer I checked how well the model predicted the winners of the US GP 2026. Project link: github.com/KumarManik44/f1-p… Live app: kumarmanik44.github.io/f1-pr… #learning #ai #datascience #statistics #dsa #aiengineer #aiengineering #harness #modelharness #modelevaluation #evals #machinelearning #python #google @GoogleOSS @Google

Another day, another great workout. 34 days out 🏃

300 Followers with a week to spare. LFG 🚀🚀

Day 27 - 180 Days of becoming a better AI Engineer After working on this for over a week, here's the live demo. #learning #ai #datascience #statistics #dsa #aiengineer #aiengineering #harness #modelharness #modelevaluation #evals #machinelearning #python #google @GoogleOSS @Google

Day 26 - 180 Days of becoming a better AI Engineer 🏎️ I rebuilt my F1 predictor from scratch. The most useful thing V4 taught me is that my V3 numbers didn't mean much. What I built: • F1 Predictor V4, a web app that gives each driver's chance of winning and finishing on the podium • You pick any remaining 2026 race, set the starting grid (or load official qualifying), and get the odds right away • Win chances add up to 100% per race and podium chances to 300%, so the numbers fit together • It runs fully in the browser with no server, and a test checks the browser and Python models agree to 12 decimal places How it's different from V3: • V3 claimed 93.89% accuracy. When I re-read the code, that number turned out to be close to meaningless • Only ~15% of drivers finish on the podium, so predicting "no podium" for everyone already scores about 85% • V3's features leaked future results: teammate stats were sorted by driver instead of by date • V3 picked and reported its model on the same small test set, and used hand-tuned multipliers for each circuit • V4 uses Brier score instead of accuracy, tests every feature for leaks, and compares everything to a simple baseline The approach: • Data from 2022–2026, with every known data error fixed by a written rule • Every feature is built only from earlier races, and a leak test rebuilds each race with all future data deleted (109 of 109 races matched) • I wrote down every keep/ship rule before seeing its result, so I couldn't move the goalposts later • 10 races were held back and opened exactly once, at the very end • Logistic regression, CatBoost and XGBoost were all tested against a plain "historical podium rate per grid slot" lookup What I found: • The starting grid carries almost all of the signal • Driver form, team strength, gap to pole and circuit history each added nothing once the grid was known • The bigger models (CatBoost, XGBoost) did worse than the simple lookup • The shipped model is a recalibrated grid lookup: slightly better on development data, tied on the holdout • The honest result is that a grid-slot lookup is very hard to beat Real-world impact: • Fans get calibrated, believable probabilities instead of a flashy accuracy number • The app openly says that the circuit you pick doesn't change the odds, rather than inventing circuit adjustments • For anyone building ML: pick the right metric, beat a simple baseline, and lock your test set away. A smaller honest number is worth more than an impressive wrong one • Next up is a live forward test: predictions are frozen after qualifying and scored after the race, starting with Austin this weekend ▶ Try it: kumarmanik44.github.io/f1-pr… 💻 Code: github.com/KumarManik44/f1-p… #learning #ai #datascience #statistics #dsa #doctordoom #doomsday #marvel #avengersdoomsday #aiengineer #aiengineering #harness #modelharness #modelevaluation #evals #machinelearning #python #google @Google @GoogleOSS

Day 25 - 180 Days of becoming a better AI Engineer 🏎️ I rebuilt my F1 race predictor from scratch, and the most useful thing it taught me was where my old model was wrong. Last year my V3 model reported 93.89% accuracy. This year I went back to that code and ran V4 as a clean restart, phase by phase, with a written plan and rules I couldn't bend. What was wrong with V3: • Only about 15% of drivers finish on the podium, so saying "no podium" for everyone is already ~85% accurate. Accuracy was the wrong metric. • Team stats were sorted by driver instead of date, so some features could see the future. • "Circuit" features were really career totals, and missing values were filled with 0. • The model was chosen and reported on the same small test set. What V4 does differently: • 📏 Brier score as the headline number, with confidence intervals that resample whole races, not individual drivers. • 🔒 A leak test that rebuilds every feature with all later races deleted. Identical in 109/109 races, and a deliberately leaky "canary" feature got caught every time. • 🧱 A simple baseline first: a smoothed podium rate per grid slot. Every model had to beat it. • 🧪 Feature ablation with a keep rule set in advance: qualifying gaps, driver form, team strength, circuit history. • 🤖 Logistic regression, CatBoost and XGBoost, plus calibration tests. • 🎯 A 10-race holdout, opened exactly once, with a lock file so I couldn't quietly try another model afterwards. The honest result: • None of the extra features improved on the starting grid. • CatBoost and XGBoost did worse than the simple lookup. • The best model, a recalibrated grid lookup, beat the baseline on development data but tied it on the holdout: Δ Brier −0.0004, 95% CI [−0.0018, +0.0011]. • Conclusion: in modern F1, the starting grid is very hard to beat. I'd rather publish that than a flattering number. What I shipped anyway: • A web app (React, TypeScript, Tailwind, Motion). Pick any remaining 2026 race, drag drivers around the starting grid or load official qualifying, and get every driver's P(win) and P(podium). • The model runs entirely in the browser, and an automated test checks it matches the Python notebook to 1e-12. • A "How good is it, honestly?" section on the page itself, showing the holdout result. What I'm taking away: • Pick the metric before you see the results. • Build the baseline before the model. • An honest "no improvement" is a result, not a failure. Code, notebook and full write-up: lnkd.in/g5inQG2Q #learning #ai #datascience #statistics #dsa #doctordoom #doomsday #marvel #avengersdoomsday #aiengineer #aiengineering #harness #modelharness #modelevaluation #evals #machinelearning #python #google @Google @GoogleOSS

Lockin season baby 37 days out 🏃🏻

Day 24 - 180 Days of becoming a better AI Engineer I built an app where Doctor Doom counts how many Instagram Reels I scroll each day. Now it's open source. Doom Scroll: - Counts every Reel you scroll, live, in a small overlay at the edge of the screen - Doctor Doom's face changes as your count climbs - Shows today's total plus your last 7 days - 100% on-device: no account, no internet, no tracking - Android only (iOS doesn't let one app see another) How it's built: - React Native + TypeScript for the app - A Kotlin Accessibility Service in the background that detects each Reel swipe - Counts stored locally on your phone Why: "I'll just watch a few" never ends at a few. Seeing the real number on screen while you scroll works better than any screen-time report you'll never open. It's early, and I'll keep improving it. Ideas, issues and PRs are welcome. 🔗 github.com/KumarManik44/doom… #learning #ai #datascience #statistics #dsa #doctordoom #doomsday #marvel #avengersdoomsday #aiengineer #aiengineering #harness #modelharness #modelevaluation #evals #machinelearning #python #google @Google

"You'll be shocked when you discover how easy it is to part ways with people forever. That's why, when you find someone you wanna keep around, you do something about it." - How I met Your Mother S09 Ep21

It all leads to Doom. Join the waitlist doomscrollcount.vercel.app/

Day 23 - 180 Days of becoming a better AI Engineer It all leads to doom 💚doomscrollcount.vercel.app/ @Himanshuraj524 Doom Scroll: An app that tracks how many Instagram Reels you scroll each day. Background counting: a Kotlin Accessibility Service detects forward reel changes, ignoring repeated events and backward swipes. Live Doom overlay: a small Doctor Doom pill displays today’s count over Instagram while Reels is open. Stats dashboard: a React Native screen shows today’s total and seven days of history with comparison bars. Setup and troubleshooting: instructions link directly to Accessibility settings and battery settings to help keep counting active. Local storage: daily counts stay on your phone in Android SharedPreferences, with a fresh count for each local calendar day. No account or backend. #learning #ai #datascience #statistics #dsa #doctordoom #doomsday #marvel #avengersdoomsday #aiengineer #aiengineering #harness #modelharness #modelevaluation #evals #machinelearning #python #google @Google