Systems engineer who builds AI agent infrastructure. I write the layers most people skip: schedulers, kernels, wire protocols, and now, provably fair runtimes for multi-agent LLM systems.
CS undergrad at MIT World Peace University, Pune (B.Tech CSE, CGPA 8.25/10, expected May 2027). I've written an operating system from the boot sector up, a lock-free key-value store that hit 49,000 ops/sec, and firmware that held a hyperloop pod at 11 ms end-to-end latency under live European competition. Most recently, I proved a fairness theorem for scheduling AI agents and built the runtime that enforces it.
Currently: open to software engineering and AI infrastructure internships.
Multi-agent LLM systems run with concurrency but no concurrency policy: one agent's 100K-token call can stall every sibling, and nothing enforces a budget until the bill arrives. I wrote a systems paper on this problem and built the runtime that fixes it.
- Designed Reconciled Virtual Time (RVT): extends virtual-time fair queueing to API calls that are non-preemptible and whose cost is unknown until completion, by charging an estimate at dispatch and reconciling the truth when the call returns
- Proved a closed-form service-gap bound (Theorem 1) and starvation freedom as a direct corollary, plus two negative results that delimit the design space: unreconciled estimation is provably unbounded, and non-mediating middleware cannot enforce fairness at all
- Modeled each LLM agent as an OS-style process: private context, budget vector, virtual time, and capabilities, all behind one mediated syscall interface (
infer,invoke,send/recv,clock,random) - Built the ~2,400-line Go runtime and a deterministic evaluation harness; measured a 0.9998 Jain fairness index (vs. 0.162 for FCFS and round-robin) under 31:1 token-cost skew, at 1,847 ns/call scheduling overhead, tested up to 50 concurrent agents
Read the paper and code → github.com/atharvaarbat/agent-kernel
AI Agent Engineer Intern, OVRT · Remote · Jan 2026 – Jun 2026 Built a TypeScript multi-agent orchestrator coordinating 6 LLM agents for Shopify automation; designed a secure Hono API proxy with HMAC signing that cut latency 20%; shipped a 24/7 monitoring agent processing thousands of events daily.
Navigation and Control Lead, Vegapod Hyperloop · Pune, India · Aug 2023 – Aug 2025 Architected the pod's C++ control unit, holding 11 ms average end-to-end latency in live runs at European Hyperloop Week 2024 (Zurich) and 2025 (Groningen). Led a 9-engineer subsystem to an 8th/16 finish.
Frontend Developer Intern, BytesWrite · Pune, India · Aug 2024 – Nov 2024 Built an internal ops platform (Next.js, TypeScript) adopted by 150+ employees; raised the Lighthouse score from 54 to 91.
| Project | Stack | What it does |
|---|---|---|
| AgentKernel | Go | Fair-scheduling kernel for multi-agent LLM systems, with a proven service-gap bound. 0.9998 Jain fairness, 1,847 ns/call overhead |
| ConcurrentKV | Go, TCP | Redis-style KV store over a custom binary protocol. 49,000 ops/sec with a lock-free MPSC queue, benchmarked against Redis |
| ARBOS | C, x86 Assembly | 32-bit OS booting on real hardware: bootloader, paging, interrupts, drivers, and a round-robin scheduler |
| Auth4Agent | Go, Ed25519, JWT | Decentralized identity protocol for AI agents; private keys never leave the agent |
| LLM Code Review Agent | Python | Agentic CLI that reviews git diffs into structured JSON. Cut token usage 35%, 1.8s p95 latency |
| System Design Mafia | TypeScript, Next.js | Live platform for learning system design through interactive diagrams: sdmafia.hdtl.in |
| Gym Management | Next.js, Prisma | Full-stack platform running live at 2 gyms in India |
More projects
| Project | Stack | Description |
|---|---|---|
| SaaS Auth Starter | Next.js, Prisma, PostgreSQL | Production auth stack: JWT sessions, WebAuthn passkeys, OAuth, email OTP |
| Track Matching System | Python | Offline AI matching engine for Army intelligence; SBERT embeddings + ANN search |
| Huffman Compressor | C | 42% average compression at ~180 MB/s, benchmarked against gzip |
| Raytracer Engine | C, SDL2 | 2D real-time raytracer with interactive light and shadows |
| Custom Shell | C | Unix-like shell built with fork()/execvp() |
| XORCrypt | C | File encryption with 256-bit XOR key generation |
| Terminal Image Renderer | C, stb_image | Renders images as colorized ASCII art in the terminal |
Languages
Systems & Infra
AI & Agents
Web
- Speaker, DevConf.in 2026: "Building Real-Time Control Software" (watch talk)
- Competed at European Hyperloop Week 2024 (Zurich) and 2025 (Groningen); placed 8th of 16 international teams
- Solved 250+ LeetCode problems: dynamic programming, graphs, trees, binary search