I like software that knows when to stop.
Voicy makes you check the person before it sends. SpaceATC stops at a human decision before it moves a satellite. VidWise puts the source moment under the answer. OmniCommand removes the file-conversion tool hunt. codex-spend shows where the tokens went.
I study Mathematics and Computing at DTU, graduating in 2027. In August 2026, maintainers merged 22 of my pull requests across t3code, Inspect Robots, Webcmd, and Agent Orchestrator. I am looking for software engineering and AI internships.
You have report.pdf. You want Markdown. You should not need to remember which parser or flag gets
you there.
npm i -g omx-cmd
omx convert report.pdf to markdownomx picks the local engine and writes the result beside the input. It handles documents, images,
audio, and video, plus batch jobs and JSON output. Version 1.1.0 has
125 tests.
The interaction is one sentence.
"message Pulkit that I'll be late"
Hold Ctrl+Space, say it, then check the contact and message on a confirmation card. Voicy sends
through WhatsApp Desktop only after you confirm. Speech and contact matching stay on the Mac.
A correct transcript sent to the wrong person is still a failure. In a 72-case evaluation, Voicy reached 90.9% top-1 recipient accuracy with zero wrong-person sends.
Codex usage is easy to feel and hard to see.
npx codex-spendThat command prints a terminal summary, then opens a local dashboard with usage by model, project, day, token type, and cache. It reads Codex session data from your machine and uploads nothing.
Two satellites are converging. Who moves?
In this four-person project, two operator agents propose burns. A coordinator compares them. The workflow stops for human approval. In the injected demo encounter, an approved 0.242 m/s burn reports a miss-distance change from 0.463 km to 3.391 km. This is a systems demo, not flight software.
I handled integration, live-testing fixes, and the encounter data layer.
Six videos. One question. Clickable evidence.
Paste up to six YouTube links and ask across all of them. VidWise searches their transcripts and attaches timestamp citations to the answer. Click one to inspect the source moment instead of scrubbing through every video.
The repository includes tests and an evaluation runner. The current baseline covers two of five retrieval configurations, and human review is still pending, so I do not publish a final accuracy number.
A lot of my work happens in repositories I did not start. In August 2026, maintainers merged 22 of my pull requests:
- 9 fixes in t3code across server, desktop, mobile, and shared code.
- 11 fixes in Inspect Robots, including evaluation math, CLI validation, logging, and controller state.
- OmniSearch for Webcmd, with commands for public search and normalized JSON.
- Dock icon bounces for Agent Orchestrator when notifications arrive on macOS.
At Zanista AI, I built crawlers and cleaning pipelines processing more than 10,000 articles a day, then added summarization, entity extraction, sentiment, and clustering.
At BharatFare, I triaged a 281-issue backlog and shipped more than 40 fixes across core user flows.
If something breaks, open an issue. I would rather get a useful bug report than a polite star.