PRINCEKUMAR

Students repeat the same search, résumé tailoring, cover-letter, and tracking work for every application.

We brought role discovery, ATS-ready application materials, and application tracking into one focused workflow.

Finch is live for internship and new-grad searches, with thousands of roles available to browse without an account.

The same event can be priced differently across two prediction markets, but the gap can disappear before a person can compare and place both orders.

I built separate live feeds for Kalshi and Polymarket, then kept the comparison and trade decision path small enough to react immediately.

The private engine watches both markets at once, checks the opportunity and risk, and sends paired orders when the prices line up.

Shopping lists are written in recall order, not aisle order, so people backtrack through the store. Existing apps locate one item but do not sequence the whole trip.

I converted messy store-guide PDFs into walkable maps, aligned live shelf locations to those maps, and only showed a store after its placement checks passed.

A 25-item pilot dropped from 818m to 426m. The route redraws on every list change, and seven stores now run through one repeatable onboarding pipeline.

When several coding agents are running, their sessions, approval prompts, and usage windows live in separate apps. That makes the simplest question—what is active?—surprisingly hard.

I used the physical notch as ambient status space and read existing local session data instead of adding another account, server, or Electron app.

One lightweight macOS panel shows live sessions, limits, multiple accounts, and optional approval controls while keeping all activity on the machine.

Small lesions disappear when a high-resolution face photo is squeezed down for a model. Strong-looking lab metrics also failed to tell me when a model broke on real pipeline images.

I ran an A/B test on native-resolution tiles versus zoomed images, kept failed experiments in the record, and added visual checks before trusting aggregate scores.

Tiling lifted recall from 4% to 70% in the serving test. The final detector reaches 0.743 recall, and product ranking uses the simpler baseline that beat the learned model.

Blocked crew hours are wasted when a sequence spoils, but the exports contained fields written after the outcome. A high score would be meaningless if the model peeked into the future.

I audited leakage, split the data by time, used only pre-departure snapshots, and reported calibrated risk so schedulers could choose their own action threshold.

The final blend reached 0.822 ranking AUC and won the TAMU ML Club competition. It also identifies severity reliably, not just whether something may go wrong.

Two teammates can ask different coding agents to change the same function in incompatible ways. The collision usually appears only at merge time, after both sides have done the work.

Synapse compares before-and-after software contracts rather than whole files. Its decision rules are deterministic; optional AI can explain a warning but cannot weaken it.

Teams get a warning before the second edit. The package works across TypeScript, JavaScript, Python, and Go, while raw code stays on each developer’s machine.