I BUILD
THINGS THAT
ACTUALLY
RUN.
I build web applications, SaaS products, and backend systems. Along the way, I explore how AI works from the ground up.“I don't just construct user interfaces — I understand what happens underneath the stack.”
BUILT. TESTED. SHIPPED.
Empirical evidence of working systems over buzzwords. Ground-up implementations, native runtimes, and verified code.
C++ / GPT-2 INFERENCE ENGINE
Built a GPT-2 inference engine from scratch in C++.
AI CODING HARNESS
Building a Claude Code-style local AI development harness.
OPEN SOURCE PROJECTS
Building and experimenting in public.
SYSTEMS ENGINEERING
Working close to runtime, infrastructure and developer tooling.
THINGS I'VE BUILT.
Engineering projects and functional systems built from first principles. Rich previews with architecture diagrams by default — expand any case study for deep dives.
GPT-2 INFERENCE ENGINE
Transformer forward pass implemented from first principles in C++17.
A GPT-2 inference engine implemented in C++ to explore the mechanics of transformer inference. The project focuses on model execution, tensor operations, transformer computation, and token generation without relying on a high-level inference runtime. Loads GPT-2 Small binary weights into contiguous memory buffers and executes the full forward computation graph. Generates text autoregressively with causal self-attention and temperature-scaled sampling.
- ■Custom tensor buffer memory management without runtime allocations during forward loop
- ■Manual matrix multiplication kernels with cache-friendly loop tiling and SIMD-friendly layout
- ■Full Transformer block: LayerNorm, causal masked MHA (12 heads), MLP with GELU, and residual sums
- ■Greedy and temperature-based autoregressive decoding loop predicting 50,257 vocabulary logits
AI CODING HARNESS
Local CLI software-engineering harness inspired by agentic coding workflows.
A local developer tool inspired by agentic coding workflows, built to experiment with model APIs, tool execution and software-engineering agents. Implements an autonomous agent loop that executes commands, inspects files, and iterates on code within terminal sessions. Provides deterministic tool dispatch with compiler diagnostics and error recovery feedback loops. Compacts and prunes conversation context to prevent prompt overflow during long debugging runs.
- ■Pluggable provider interface supporting Anthropic Claude 3.5 Sonnet and OpenAI GPT-4o
- ■Deterministic tool calling protocol: bash runner, file read/write, pattern grep, ripgrep inspection
- ■Conversation context compaction and pruning to prevent prompt overflow during long debugging runs
- ■Robust error recovery: feedback loops when shell commands fail or edits do not match file markers
CEX MATCHING ENGINE & ORDER BOOK
First-principles centralized crypto exchange matching engine and transaction ledger.
A centralized crypto exchange (CEX) core built from scratch to demystify high-throughput financial infrastructure. Implements an in-memory limit order book matching engine with FIFO price-time priority, double-entry balance locks, atomic trade executions, and real-time order book broadcasts. Decouples fast in-memory execution from asynchronous ledger persistence for auditability and sub-millisecond matching performance.
- ■In-memory price-level red-black tree / sorted doubly linked list order book for deterministic sub-millisecond matching
- ■Strict price-time priority matching algorithm for both LIMIT and MARKET orders with partial fills
- ■Double-entry bookkeeping preventing fractional balance discrepancies or double-spend race conditions
- ■Event-driven transaction log for state reconstruction and failure recovery
TOOLS & TECHNOLOGIES.
Technologies I use to build full-stack web products, plus systems and AI tools I explore from first principles. Categorised by real-world usage.
FRONTEND & WEB APPS
Crafting responsive web applications, interactive interfaces, and design systems with strict TypeScript types, server components, and clean CSS.
BACKEND & APIS
Designing HTTP/RESTful services, event-driven WebSocket connections, background processes, and exploring Go and Python backends.
DATABASES & DATA
Relational schema modeling, type-safe queries, in-memory caching, double-entry financial ledgers, and data persistence guarantees.
AI & MACHINE LEARNING
Exploring model architectures from the ground up: Transformer attention mechanisms, local CLI agent loops, tool dispatch protocols, and on-device inference.
SYSTEMS & PERFORMANCE
Investigating performance close to hardware: cache-conscious loops, zero-allocation buffers, SIMD parallelism, and memory efficiency.
DEVOPS & INFRASTRUCTURE
Operating inside Linux environments, containerizing services with Docker, git-centric continuous integration, and cloud service deployments.
I LIKE KNOWING
WHAT'S HAPPENING
UNDERNEATH.
From model inference and memory to APIs, containers, Linux and cloud infrastructure — I enjoy understanding the systems behind the interface.
SOURCE & SYNTAX
Strictly typed, idiomatic code in C++, TypeScript, Go, and Python. Designing explicit contracts, modular abstractions, and zero-redundancy interfaces.
WHAT I'M WORKING ON NOW.
Active projects, ongoing experiments, and engineering initiatives currently in progress.
A developer-tooling project exploring AI-assisted coding workflows, multi-provider model integration, deterministic tool execution loops, and automated feedback diffs.
An engineering investigation exploring local AI model inference inside browser and Chromium runtimes for private DOM reasoning and zero-cloud API workflows.
A transformer forward pass implemented from first principles in standard C++17, focused on understanding the mechanics of GPT-2 inference without PyTorch, ONNX, or runtime libraries.
I BUILD.
I BREAK.
I UNDERSTAND.
I SHIP.
LOCATION: Pune, India [UTC+5:30]
PRIMARY FOCUS: FULL-STACK WEB & SAAS
EXPLORING: AI MECHANICS & SYSTEMS INFRASTRUCTURE
DIRECT MANIFESTO
I'm Vinayak Dhemare, a full-stack software engineer focused on building web applications, SaaS products, and reliable backend infrastructure.
My primary focus is building full-stack software that is useful, reliable, and user-centric. Alongside product development, I explore how modern systems work from first principles — understanding the mechanics underneath APIs, runtime engines, and language models.
Whether architecting real-time streaming interfaces, writing transactional ledgers, or deconstructing GPT-2 inference in pure C++ — I care about deterministic behavior, clean architecture, and shipping things that actually execute.
FIRST PRINCIPLES.
Four guiding invariants for building systems that last.
UNDERSTAND THE SYSTEM
“Don't just make it work. Understand why it works.”
Abstractions are convenient until they leak. Peeling back high-level libraries to see memory buffers, network packets, and CPU instructions builds true engineering intuition.
BUILD BEFORE OVERENGINEERING
“Start small. Measure. Then make it better.”
Avoid solving problems you do not have yet. Write working code, profile the hot paths under realistic workloads, and optimize with empirical benchmark data.
PERFORMANCE IS A FEATURE
“Latency, memory, cost and reliability matter.”
Speed is not an afterthought. Thoughtful data structures, cache-conscious memory layouts, and low overhead translate directly into user responsiveness and lower cloud costs.
SHIP THE THING
“A finished imperfect system teaches more than an unfinished perfect architecture.”
Architectures drawn in diagrams prove nothing until bytes execute on real silicon. Build the prototype, run the tests, deploy it, break it, and iterate in public.
HAVE SOMETHING
WORTH BUILDING?
Have a project, idea, internship opportunity, collaboration or interesting engineering problem?
Let's talk.