Software engineer · Product, AI agents & systems
Waterloo dropout. Founding engineer. Open-source contributor.
Built ReferRush from its first commit. Early engineer at extraordinary.com. Now building Rust agent infrastructure, with merged contributions to Atuin, Podman Desktop and 11 other repositories.
My changes are merged in Atuin, Podman Desktop and 11 other public repositories. Each credit links to the accepted work.
Hook response compatibility
UI error handling, file inputs and extension organization
License contribution
Authentication fix
Unread-message trackbar and recovery-key verification
Worker IDs in task-spawn telemetry
All merged external PRs ↗ · Repository stars and merged counts checked October 2, 2026.
Feb 2025 - Sep 2026
Founding Engineer · Bengaluru
Jan 2024 - Jan 2025
Backend Engineer
Oct 2022 - Nov 2024
Founder & Software Engineer · Remote (Dubai)
Oct 2022 - Sep 2024
Software Engineer
Rust, Tokio, Axum, redb, MCP
Built local-first agent memory in Rust: atomic topic/tag/time indexes, bounded BM25 search, revisioned writes and native OpenCode MCP. Measured 95.13% Recall@10 across all 500 LongMemEval-S questions with 14.34 MiB sampled server RAM on an M4 Pro.
Full LongMemEval-S answer accuracy: 85.20% with GPT-6 Luna as reader and judge, a model variant rather than official leaderboard parity. The memory engine needs no model or cloud calls. Source MVP; phone bindings and semantic search remain planned.
Rust agent runtime
Built remote execution that continues after client disconnects, streamed tool output, persisted sessions, event replay and crash recovery. A separate side agent reports progress without interrupting active work.
Verified session recovery after a container restart on a VPS.
TypeScript Firebase SDK & CLI
Built typed Functions/Firestore helpers, generated configuration and staged rebuilds. Verified HTTP, Hosting, callable functions and Firestore triggers in real emulators; added Node.js 22/24 CI.
Python computer-use testing
Built Linux desktop testing through accessibility, screenshots, video and reports. Added authenticated APIs, commit-pinned jobs, atomic queue claims, webhook deduplication and restart recovery.
Fixed host seccomp incompatibility blocking desktop startup.
Rust HTTP/1.1, Tokio, SQLite WAL
Built a raw HTTP service with commit-before-ack writes, live backups and rollback. Measured 1.34M redirects/s on a four-vCPU VPS: loopback, pipeline depth 512, median of three eight-second runs, zero errors.
The same comparison measured 34,423 redirects/s without pipelining. These are controlled local measurements, not public HTTPS capacity.
Rust terminal AI assistant
Built streamed model/tool output, a live observer feed, bearer authentication and process-group timeouts; deployed verified commits through CI with rollback.
Also built Plain Corners, a native macOS utility, and this Windows XP portfolio in React, Next.js and TypeScript.
Current instantKV retrieval runs: Apple M4 Pro, 24 GiB RAM, macOS 27. Server RAM is sampled Rust process RSS; query time excludes model inference.
| Benchmark | Recall@10 | Server RAM / query p95 |
|---|---|---|
| LongMemEval-S · 500 questions | 95.13% | 14.34 MiB / 5.65 ms |
| LoCoMo · 1,533 scored questions | 57.66% | 11.25 MiB / 0.79 ms |
| BEIR SciFact · 300 questions | 81.43% | 23.14 MiB / 1.55 ms |
| BEIR ArguAna · 1,406 questions | 76.96% | 24.78 MiB / 46.83 ms |
| BEIR NFCorpus · 323 questions | 15.31% | 22.92 MiB / 0.93 ms |
LongMemEval-S answer accuracy: 85.20% (426/500), 95% confidence interval 82.0–88.2%. GPT-6 Luna was reader and judge; this is a model variant, not official leaderboard parity. ArguAna had 688 truncated and 1,149 reduced queries. Timings include shared host load. All run conditions, limits, comparisons and raw results ↗
Separate structured-memory test: three fresh 10,000-record databases, 512-byte content plus metadata. Recorded binary: 8.31 MiB. All 30,000 memories recovered after process restarts; device power loss was not tested. Performance report ↗
Earlier HTTP throughput runs: shared four-vCPU VPS, client and server on loopback.
| Workload | Result | Conditions |
|---|---|---|
| instantKV knowledge GET | 4,010 req/s | 16 concurrent clients; p99 47.20 ms |
| instantKV durable PUT | 744 req/s | 8 concurrent clients; immediate commits; p99 57.65 ms |
| rushort redirects | 1.34M req/s | 32 connections; pipeline depth 512 |
| rushort round trip | 34,423 req/s | 64 connections; no pipelining |
Median of three runs per workload, zero errors. instantKV completed 157,500 measured requests across five profiles. These runs measure specific workloads, not public HTTPS capacity.
University of Waterloo dropout | Computer Engineering studies | Left to build full time
Hacked rice.edu at 17 and responsibly disclosed SQL injection. Hacked into Waterloo’s GPU cluster. 3x National Math Olympiad silver (ages 15–17); IITM NexHack 2nd (2023); IPSC Web Dev 4th in India.