awoken.in LinkedIn Email X Pull requests
CS student at Newton School of Technology. I take ideas from a rough sketch to something people can run, then check whether it actually works. Recent work has taken me from a C++ camera-control loop to AI agents that pay for outcomes over x402 to Rust terminal internals.
Source: margin402, FSOC, gec-platform. FSOC was built with Team IRODOV for ISRO's SIH26169 problem statement; its numbers are recorded simulation results, and the repo documents the tradeoffs. Margin402's provider market is simulated; its payments are real Algorand Testnet transactions.
What changed in each pull request
HADES CLI #32 (merged). The TUI enabled global mouse capture at startup even though every interaction is keyboard-driven, so terminals sent drags to the app instead of selecting text. Removed the capture on init and kept the cleanup on exit. Ratatui, Crossterm.
HADES CLI #30 (merged). Added /export and /import. Exports Markdown or JSON; imports HADES exports, ChatGPT conversations.json, Claude transcripts and generic Markdown with format detection, then restores them into the active context. Spans the storage, core and TUI crates.
awesome-ai-apps #317 (merged). A FastAPI service paired a wildcard CORS origin with allow_credentials=True. Origins now come from an explicit allow-list that fails closed when unset, with regression tests for the edge cases raised in review.
Node.js #66376 (open). Backports two V8 fixes to v24.x-staging: a wrapper already marked as dying could be re-added to a WasmCodeRefScope, tripping a JIT allocation CHECK. An independent tester reported 0 crashes in 191 repro runs with the patch, versus 8 in 149 without.
LLMVault #44 (open). Per-player, per-challenge attempt tracking, a 30-second cooldown after five misses, and HTTP 429 with retry_after. 19 tests.
| Project | What it is | Stack |
|---|---|---|
| ChallanCheck | Checks whether a traffic e-challan's photo actually shows the cited violation. A vision API observes; deterministic rules decide. | JavaScript, AWS Lambda, Rekognition |
| Beacontra | Ranks marketplace listings for brand review by combining price, seller and reverse-image evidence. | TypeScript, SerpApi |
| Intervu | Adaptive interview engine: each answer updates a model of what the candidate knows and picks the next question. 48-hour team build. | Next.js, TypeScript, Supabase |
| ring-escape | Arcade game: rotate concentric rings to guide a ball out of the maze. | Python, Pygame |
| Portfolio | Personal site. | React, Vite, Framer Motion |
| FolderPrettifier | Sorts a messy folder by file type without overwriting duplicates. | Python |
| Airline Reservation System | Terminal booking system with user and admin roles. School group project. | Python, MySQL |
That's why FSOC publishes its counterexamples next to its best numbers, Margin402 has a "what is real vs. simulated" table, and my pull requests start with the root cause.
