I build web apps, APIs and small tools, from a tech news feed to a local video converter.
I work mostly in TypeScript, React, Node.js and Python. I also send fixes upstream and write about what I learn.
| Project | What it does |
|---|---|
| IMA | Pulls the latest tech news into one feed, so you can keep up without the FOMO. |
| Voino | Turns meetings into editable, interactive notes and a visual board. Prototype. |
| Saartheye | A prototype I built with the vision of helping blind people use their phones to navigate. Not a validated navigation aid. |
| FacCheck.ai | A predictive-maintenance dashboard prototype for machine health and alerts. Uses CSV replay and simulated metrics, not real factory results. |
| TaskFlow AI | Helps people arrange tasks and projects, collaborate with others and use AI for planning. |
| mov2mp4 | Helps editors convert MOV videos to MP4 for free, on their own machine. |
10 merged PRs and 5 open ones in other people's projects. Open does not mean accepted. The full list with links and current status is in OPEN-SOURCE.md.
- pgAdmin - fixed pgpass handling in the Change Server Password dialog, with regression tests, and made the Database properties panel show real default privileges. Also fixed multi-byte characters split across Dashboard log reads and moved the return-set switch beside the return type. Also preserved explicit 24-hour mode in time picker popups, with regression tests and two maintainer refinements. All five were squashed into master by the maintainer. A sixth, preserving opclass options and all storage parameters in reverse engineered index SQL, is open and awaiting review, as is one adapting the OpenAI LLM provider to OpenAI-compatible servers and reasoning models.
- Notify-Chain - four merged PRs, including database reconnection with bounded backoff and API contract tests with CI coverage.
- ZecHub wiki - made the keyboard work in the 404 page search, with tests. Merged.
Tripwire checks text going into and out of LLM apps for prompt injection, secrets and personal data, with plain rules and an optional Gemma judge.
Tripwire catches the override, not the intent - a rules-only failure analysis of indirect prompt-injection screening. Technical report on Zenodo, not peer reviewed. It measures text screening, not whether an agent was protected from an attack.
I write about the parts that took longer than expected, what broke and what I'd change: dev.to/sansk_ya.
One place to start: what I learned shipping a tiny FFmpeg desktop app.