Suryansh Chaudhary — Full-Stack AI Engineer and creator of MacGet

What he works on now

His day job is agent architecture. At Spiral Kite Labs he owns the whole vertical for a multi-agent financial research desk: LangGraph for the control plane, LangChain for tool integrations, a FastAPI backend, Google Cloud Run for infrastructure, and the React frontend that makes the thing usable. The unglamorous half of that job is making the output defensible rather than merely fluent — in finance, a confident wrong number costs you the user permanently.

That constraint drives most of the architecture. Retrieval runs on Vertex AI's managed multimodal RAG Engine, because the number an analyst needs usually lives in a bar chart on page 34 rather than in a paragraph a text-only parser can reach. Safety runs as layered gates: deterministic hard gates for the things that must never depend on a model's judgement, and an LLM-as-judge soft gate for everything ambiguous. Agents get explicit state, small nodes, and narrow tool boundaries — and the parts of the system that shouldn't be a model's decision simply aren't.

MacGet — an open-source macOS download manager

MacGet is the project he is best known for, and it is free and open source. It is a SwiftUI app built on an actor-based download engine that opens chunked parallel HTTP connections, captures downloads straight from the browser over Native Messaging, and bundles yt-dlp so grabbing a video is one click rather than a terminal session. Version 1.3.0 added BitTorrent and OPDS book catalogs — and deleted the “adaptive concurrency” ramp he had shipped as a headline feature two releases earlier, because the heuristic misread a saturated link as a failed experiment and made MacGet slower exactly where it should have been fastest.

MacGet is now on Product Hunt, and the source lives on GitHub.

Oratora — live audience engagement

Oratora is live audience engagement for talks, classes and events: polls, word clouds, Q&A and quizzes that a room joins from a phone browser with a short code — no app, no account, no signup wall between a speaker and their audience. Architecturally it is a Next.js modular monolith paired with a separately-scaled Socket.IO and Redis realtime tier, load-tested at 1,000 simultaneous joins, because the failure mode that actually matters is everyone in a lecture hall connecting in the same ten seconds.

Signal Boy — WiFi survey in one command

Signal Boy started as a personal irritation: finding the dead zone in a flat required downloading an app, making an account, or paying for a heatmap licence. It runs with npx signalboy and shows a live dBm readout on a Game Boy-style LCD — a Kalman filter settles the reading while an EWMA keeps the sparkline moving. It has zero dependencies, per-OS backends for macOS, Windows and Linux, and a native Android app that reads the phone's own radio.

How he works, and writes about it

He builds in public. The build log on this site runs two to four posts a month covering LangGraph agent architecture, prompt-injection defence, agentic RAG on Vertex AI, Swift concurrency inside MacGet, and the AI-assisted development workflow he actually uses day to day — including the decisions that turned out to be wrong. Being the AI, DevOps, backend and frontend person at a startup means you stop thinking in layers and start thinking in outcomes, and the writing reflects that.

Background

Suryansh studied at Poornima Group and has spent the years since working across the full stack — Python, FastAPI and Django on the backend; React, Next.js and TypeScript on the frontend; PostgreSQL, MongoDB and Redis for data; Docker, Google Cloud Run and GitHub Actions for shipping. The AI layer on top is LangGraph and LangChain with Gemini and OpenAI models, retrieval-augmented generation, and cost-aware routing between models. The full breakdown is on the skills page.

Get in touch

He is open to collaborations, freelance work on AI agent systems, and bug reports on any of the projects above. The fastest routes are email, X, LinkedIn, or GitHub.