Systems programmer who builds AI agents. Hyperloop control firmware at 11 ms E2E · a bootable x86 OS · a 49k ops/sec KV store · multi-agent orchestration in production.
CS undergrad at MIT-WPU, Pune (B.Tech CSE, May 2027 — CGPA 8.25). I like the layers most people avoid: I've written an operating system from the boot sector up, a lock-free KV store that speaks its own binary protocol, and firmware that held a hyperloop pod at 11 ms end-to-end latency under live European competition. Lately I build the other extreme too — multi-agent LLM systems running 24/7 in production.
The three GRUB entries above aren't a joke — they're my three actual resumes. Same kernel underneath.
Currently: open to internships and interesting systems / AI-agent work.
Active: Jan 2026 – Jun 2026 · remote · ovrt.in
- Built a TypeScript multi-agent orchestrator coordinating 6 AI agents for Shopify automation — BullMQ queues, LangFuse tracing
- Designed a secure Hono proxy with HMAC auth; cut latency 20% via connection pooling; integrated OpenAI, Sarvam AI & the Vercel AI SDK
- Shipped a 24/7 monitoring agent with retry + exponential backoff, processing thousands of events daily
Active: Aug 2023 – Aug 2025 · Pune, India · EHW Zurich '24 + Groningen '25
- Architected MVCU (Master Vehicle Control Unit) in C++ — multi-process, multi-threaded TCP/UART system managing real-time pod state across 4 threads at 11 ms average E2E latency under live competition
- Designed a custom binary wire protocol over UART/TCP — heartbeat, CRC16 error detection, auto-reconnect; improved pod stability 40% YoY
- Led 9 engineers; team placed 8th / 16 globally; owned the technical roadmap and signed off on every C++ code review
Active: Aug 2024 – Nov 2024 · Pune, India
- Built an internal ops platform (Next.js + TypeScript) from scratch; shipped to 150+ employees in 3 months, replacing manual spreadsheet workflows
- Lighthouse 54 → 91 via SSR with ISR and route-level code splitting; load time ~4.2 s → ~2.5 s on a throttled 3G profile
- Decomposed a monolithic UI into 35+ reusable TypeScript components, cutting duplication ~30%
Module Size Used by
auth4agent ed25519 autonomous agents with trust issues
concurrent_kv 49k ops/s people who benchmark against Redis for fun
arbos 32-bit real hardware — the fans spin and everything
llm_code_review 1.8s p95 every diff that thought it was clean
huffman 180 MB/s files that weigh too much
gym_mgmt 2 gyms paying customers in India (both of them)
auth4agent — decentralized identity & auth for AI agents Go · Ed25519 · JWT
Machine-native authentication protocol for autonomous agents: DID generation, Ed25519 keypair management, and a challenge-response proof system where private keys never leave the agent — eliminating static API key risk. Scoped JWT issuance with JWKS, replay protection (nonce- and expiry-bound challenges), offline verification, deny-by-default operator policy. Full Cobra CLI (init → register → issue → revoke → whoami) plus a companion auth server with DNS-based domain verification.
concurrent_kv — multi-threaded key-value store Go · TCP
Redis-inspired store speaking a custom binary TCP protocol: ~49,000 ops/sec, 8-worker thread pool fed by a lock-free MPSC queue, O(1) LRU eviction (doubly linked list + hash map), WAL-based crash recovery. Benchmarked head-to-head with Redis via redis-benchmark.
arbos — the reason this page boots C · x86 Assembly
A 32-bit x86 operating system running on real hardware and QEMU: multiboot-compliant bootloader, bitmap physical memory manager, 4 KB paging, IDT + PIC IRQ routing, PS/2 keyboard and VGA text-mode drivers over memory-mapped I/O, round-robin scheduler, minimal syscall interface.
llm_code_review — agentic code review CLI Python
Pipes git diff through an LLM into structured JSON reviews {severity, line_range, issue, fix} across all languages. Hunk-level context chunking cuts token usage ~35% vs. naive full-file prompting; retry + exponential backoff; ~1.8 s p95 on diffs ≤ 500 LOC.
huffman — compression engine C
42% average compression on text at ~180 MB/s: min-heap priority queue from scratch, canonical Huffman encoding for deterministic serialization, correct EOF padding on arbitrary binaries. Benchmarked against gzip.
gym_mgmt — gym management platform Next.js · TypeScript · Prisma
Full-stack platform sold to and running at 2 gyms in India: member & attendance tracking, payments, diet & workout plans, revenue analytics dashboard.
$ ls /opt — more in the archive
| Module | Stack | Description |
|---|---|---|
| SaaS Auth Template | Next.js, Prisma | Production-ready auth starter — OAuth, passkeys, sessions |
| Track Matching System | Python | Offline AI matching engine for Army intelligence; SBERT embeddings + ANN search (hnswlib) |
| Raytracer Engine | C, SDL2 | 2D real-time raytracing; ray-circle intersection, interactive light, shadows |
| Custom Shell | C | Unix-like shell with fork()/execvp(), parser, built-ins |
| XORCrypt | C | File encryption with 256-bit XOR key generation; any file type |
| Terminal Image Renderer | C, stb_image | Images → colorized ASCII art in the terminal |
$ lspci -vv — full device tree
00:01.0 — Languages
00:02.0 — Systems & Tools
00:03.0 — AI & Agents
00:04.0 — Web & Data
[ OK ] Started telemetryd — live panels below update themselves. Zero maintenance.
$ dmesg | grep -i panic
[ 404.000000] Kernel panic — not syncing: attention divided by zero
[ 404.000001] Recovery: coffee.reload() && git push --force-with-lease
[ 404.000002] System recovered. It was probably a race condition.
$ sudo shutdown -h now
[ OK ] Stopped ATHARVA OS.
[ OK ] Reached target Power-Off.
It is now safe to hire the developer.