Systems engineer & founder · Bogura, Bangladesh
I build execution engines for AI, protocols for agents, and a computer architecture in my native language — from the instruction set up.
_{9th-grade student and self-taught engineer. Everything below was designed and built at age 15. I'm 16 now, and still building.}

Now: Co-founder, Lead Engineer & CTO at Truck Pai · Founder & Engineer at Brud Code · Founder & Systems Researcher at Akkhar-Labs

Open to: research collaborations, fellowships, and ambitious projects in systems engineering, AI infrastructure, and agent protocols.

• 🚚 ~1,500 active users & drivers on Truck Pai, a production logistics platform I built and run.
• ⚙️ Brud Code: on the VS Code Marketplace, on GitHub; patched 467 files with 0 failures on Truck Pai's production codebase.
• 📄 Published research: HashedContext (up to 88.7% less GPU KV-cache per agent and 3.6× more agent turns per hour in simulation) and ZRCP, both archived on Zenodo with permanent DOIs.
• 🧾 Co-author of an accepted protocol extension: ASM Trust Delta Receipt v0.1. My design additions were merged and are cited in the ASM paper.
• 🧠 Endorsed by a Google DeepMind researcher (founding team, Gemini Deep Research) for Aura-Arc (source).
• 🔣 Designed a 16-bit CPU instruction set in Bengali, with my own assembler and simulator that run it end to end.

Repo · VS Code Marketplace · Website · MIT

Brud Code lets any AI chatbot safely edit real codebases. Every change is reviewed, logged, and reversible.

My most actively developed project: live on the VS Code Marketplace and proven on a real production codebase. It is architected as a layered, out-of-process system:

• Out-of-process daemon: the VS Code extension now runs execution and history through a standalone daemon over a typed JSON-RPC protocol (wire contracts, methods, error definitions), with a lifecycle supervisor, a client harness, a daemon staging/packaging pipeline, and end-to-end tests.
• Microkernel with a plugin system: @brud/kernel has a Level-1 PluginManager with permission contexts, topological dependency ordering, fail-fast rollback, dynamic lifecycle, and unforgeable official-plugin branding with namespace protection.
• Layered services on the kernel: a SettingsService (atomic storage, safety lattice), a WorkspaceService (boundary containment) and a BrudFileSystem (Level-0 storage drivers, hardened PathValidator, durability hardening), all bootstrapped through createStandardKernel with injectable storage drivers.
• Universal History: a SQLite-backed session ledger with a reversibility model and a state machine, replacing the earlier snapshot-only approach.
• Reliability: a read-pipeline governor with a single-materialization envelope, single-flight execution lifecycle, hardened terminal streaming, and a repo-wide zero-any, zero-lint-error baseline.
• Monorepo (core · protocol · UI · daemon · VS Code adapter), CI on GitHub Actions, provider parity tests.
• Benchmarked on Truck Pai's production codebase (384K LOC at the time): 1,000 files created in under 1 s · 467 files patched with 0 failures · 1,000 files reverted and restored in milliseconds

Brud Prime, an autonomous multi-agent engineering system, is in active development and coming soon (announcement). _{Lineage: Akkhar Code Patcher → Sover → Brud Code.}

Org · Project Ekushey

Akkhar-Labs is my independent systems research lab. It isn't a product company. It's where I work at the lowest layers of computing. Its long-term mission, Project Ekushey (named after the 1952 Language Movement), is a complete computing stack designed in Bengali, from the instruction set to the operating system.

The bottom of that stack already exists and runs:

 .বয়ন source  ──►  Boyon (assembler)  ──►  .ব machine code  ──►  Shimu (simulator)  ──►  execution report
                         └──────────── Nirdeshika SDK: one command, full pipeline ────────────┘
                                   all targeting the Nirdeshika-16 ISA

I also defined Akkhar-Hex, a Bengali hexadecimal notation (০–৯, ক–চ). Next on the roadmap: a Bengali character encoding, 32/64-bit ISAs, the Kotha programming language, the BAPS kernel, and Matrika OS.

truckpai.com · Co-founder, Lead Engineer & CTO

I designed, built, and continuously develop the entire technical platform solo, as a single proprietary monorepo: 7 Flutter apps, 25+ Node.js modules, a Go microservice, and PostgreSQL. It runs in production, is upgraded daily, and serves ~1,500 active users and drivers. It's my largest project.

Repo · Project page · lablab.ai Agentic Economy on Arc

A deterministic C++17 core compiled to WebAssembly, Circle USDC nanopayments on Arc, a 3-way escrow handshake, and Supabase-backed crash recovery, with a Next.js 15 front end.

"Wild execution from a solo 16-year-old: deterministic C++17 core compiled to WebAssembly for financial atomicity off the JS main thread, plus a 3-way ASE handshake for trust-gap escrow and a Wasm/Supabase rehydration story for crash recovery. The unit-economics framing ($0.005 task vs $0.15 gas) is exactly the Arc pitch. Impressive." — Naman Goyal, Researcher at Google DeepMind · source

_{(I was actually 15 when I built it.)}

Professional freelancer since 2025, with ongoing client work for businesses in Bangladesh. Selected projects:

• iLab BD: training platform for mobile-repair and technology courses. I built the entire mobile app, shipped many new back-end features, and contributed across the web front end.
• My Pharma: Front-end developer (Next.js).

Formal publications, archived on Zenodo with permanent DOIs.

Paper (latest version) · DOI: 10.5281/zenodo.23668957 · Code

Autonomous coding agents re-send unchanged source files to the LLM on every step, so prompts grow with every turn. That saturates GPU KV-cache memory and limits multi-tenant serving concurrency. HashedContext is a content-addressed context architecture that removes the redundancy: each file is tracked by its SHA-256 hash, past copies become compact [HASH_STUB] references, and every version stays in a local vault for lossless recovery.

It is a dual-mode architecture:

1. Mode A (Authoritative Tail): one complete copy of each current file at the end of the prompt, for the lowest memory use
2. Mode B (Lazy Tail): history is never rewritten and only changed lines are sent, so a prefix cache can reuse almost the entire prompt

Both modes run either edge-natively inside client editors or transparently at server-side inference gateways.

Results: in a simulated evaluation with 20,000 coding agents on real open-source repositories, up to 88.7% less KV-cache per agent and, on a 16-GPU cluster serving 512 agents, up to 3.6× more agent turns per hour. With a warm prefix cache, Mode B processes 88.1% fewer prompt tokens.

Paper (Zenodo) · DOI: 10.5281/zenodo.21955969 · Spec & code

A protocol specification for removing redundant context between AI agents using content hashes.

• 🎓 Just-In-Time Learning Educational Model: proposal, draft v0.1. An education model in which a learner's own goal decides what they learn next, with guidance when they need it and autonomy when they've earned it. It includes an admissions approach, an assessment framework, and a plan for testing whether it works. "Who can't find their path, teach them. Whoever can, let them be free."
• 📖 Orchestrator of Reality: book in progress. Its multi-model critique workflow parallels recent research on debate-based oversight (e.g., arXiv:2605.27483).

Open to research collaborations, fellowships, and ambitious systems and AI-infrastructure projects.
rahatarch@akkharlabs.com