VS
BengaluruOpen to interesting conversations, talks and collaborations

Vatsal Saglani builds AI agents that test software.

I lead AI engineering at QyrusAI, where agents test software across web, mobile, desktop and APIs. My work is the systems around those agents: the execution harness and shared SDK, distributed runtimes with durable evidence, memory, and evaluation. I still work directly on models and code. I write about agent architectures and MCP, and keep a steady stream of open-source experiments going.

See the workGet in touch
Years in applied ML
7+
GitHub stars
450+
Public repos
110+
Paper citations
40+
  • Leading AI engineering at QyrusAI
  • Agents testing software on real devices
  • Building Tinker, a BYOK coding agent for VS Code
  • Writing about MCP and agent architectures
  • Shipping products with coding agents
  • Running agent swarms on local LLMs
  • Based in Bengaluru

01Systems · QyrusAI

Agents that test software, and what makes them reliable

The professional work. Each entry says what the problem was, what I did, how it works and how mature it is. Internal product names are translated into what the system actually does.

  • The problem

    Every AI feature was becoming its own wrapper around a model API, with its own retries, tool plumbing and logging.

    shipped

    Approach

    A reusable Python foundation adopted across internal AI applications. Teams define tools and instructions; the SDK runs the agent loop, routes tool calls, handles model invocation, retries and guardrails, and emits token, latency, error and trace telemetry by configuration.

    What it does

    • Standard agent loop
    • Tool routing
    • Retries and guardrails
    • Telemetry as a switch
    PythonAgent runtimeObservability

Maturity labels are conservative: shipped means in use inside the product, architected means designed and built without adoption claims, prototype means research. No metrics are quoted that I cannot stand behind.

02Open source

Things I have shipped

Side projects, mostly things I wanted for myself and could not find. The weekend builds are exactly that: some took half a day, some took several weekends, all of them were built alongside coding agents.

Developer tools · 2026 · JavaScript

Tinker

An open-source, bring-your-own-key AI coding assistant for VS Code.

A coding agent for VS Code with a pluggable provider layer (OpenAI, Anthropic, Gemini, Azure, Bedrock) and keys kept in editor secret storage. Chat-first UX with file and symbol context, tool use for code search, files, git and diagnostics, MCP servers, and affordances for reviewing and applying changes.

  • VS Code API
  • MCP
  • Multi-provider
Weekend buildActive
Repo

Agentic experiments · 2026 · Swift

Cider

A calmer corner of your Mac.

A menu-bar companion that lives in the notch: watch your coding agents work, track tasks and linked plans, keep local Markdown notes, control music and see account usage. I wanted one place to glance at while agents ran, so I built it over a few weekends with the agents themselves.

  • Swift 6
  • SwiftUI
  • macOS 26
1Weekend buildActive
Repo

Agents & LLMs · 2024 · Python

GraphRAG4Rec

GraphRAG, from scratch, for movie recommendations.

A naive implementation of Microsoft's GraphRAG over the IMDB Top 1000: build the entity graph, summarise communities, then answer recommendation queries with graph-aware retrieval.

  • Python
  • LLMs
  • Knowledge graphs
74Stable
Repo

Agents & LLMs · 2024 · Python

claudetools

Function calling for the Claude 3 family, before it was built in.

A Python library that gave Claude 3 models OpenAI-style tools and function calling, including vision inputs. Published on PyPI and used as a drop-in replacement in existing agent code.

  • Python
  • Anthropic API
  • PyPI
40Stable
Repo

Agentic experiments · 2026 · Swift

Capsule

Compose-style orchestration for Apple's container runtime.

A native container manager for Apple silicon: a SwiftUI app plus a capsule CLI sharing one engine that translates Compose files into native volumes, networks and port publishing. Built because I wanted Compose-style workflows on Apple's own container runtime. Honest about the subset it supports.

  • Swift 6.2
  • SwiftUI
  • SPM
  • Apache-2.0
1Weekend buildDeveloper preview
Repo

Agentic experiments · 2026 · Swift

rafu

A native repo editor for the age of coding agents.

Not an IDE. A focused macOS editor for the targeted edits you make after a terminal agent has done the heavy lifting: restorable workspaces, TextKit editor groups, git changes, history and side-by-side diffs. A weekend build that filled a gap in my own workflow.

  • Swift 6.2
  • SwiftUI
  • TextKit
  • MIT
Weekend buildIn progress
Repo

Agents & LLMs · 2024 · Python

llama-cpp-search

A GenAI search engine that runs on a laptop.

Web search, scraping and answer synthesis powered by llama-cpp-python and Phi-3, showing how far a small language model goes with the right chain.

  • llama.cpp
  • Phi-3
  • Playwright
43Stable
Repo

Agents & LLMs · 2025 · Python

swarmloka

Local agent swarms with Qwen 2.5.

An OpenAI-Swarm-style framework for orchestrating hand-off based agent swarms entirely against local LLM servers (LM Studio, llama.cpp, Ollama). Grew out of the local-qwen-swarm experiment.

  • Qwen 2.5
  • Async Python
  • Local LLMs
15Experimental
Repo

All projects · 18

All repositories on GitHub

03Experience

Seven years of applied ML

From classical NLP and computer vision to transformers, and now agent systems in production.

Jul 2024 — Present

Data Science Lead, GenAI

QyrusAI · BengaluruQuinnox, Jul 2024 – Feb 2026 · Qyrus India, Mar 2026 – present (one continuous role)

Lead AI engineering for QyrusAI: agents that test software across web, mobile, desktop and APIs, and the systems that make them reliable.

  • Built and maintain a shared Python agent SDK and execution harness adopted across internal AI applications, standardizing model calls, tool execution, retries and guardrails, with token, latency, error and trace telemetry.
  • Led Device Farm AI Sessions: users state a testing objective in plain language and agents operate real Android and iOS devices to produce reusable tests, with live guidance, parallel sessions, reruns and step-by-step evidence.
  • Architected execution and reporting on AWS AgentCore and ECS, separating live session streams from durable run history and artifacts, with CI/CD integration and enterprise deployment support.
  • Python
  • Agent runtimes
  • AWS AgentCore
  • ECS
  • Memory and retrieval
  • Evaluation
  • Playwright
  • PyTorch

Nov 2021 — Jul 2024

Senior Data Science Engineer

Quinnox (QyrusAI product) · Bengaluru

Took QyrusAI from classical ML models to generative AI, building the first LLM-powered testing products.

  • Led development of the Generative AI Software Testing Kit (STK), a Python package that let traditional QA teams use AI in their existing workflows.
  • Designed a cloud API-testing agent on LLMs with robust fallback strategies, and copilot bots for web, mobile, API and desktop testing.
  • Trained transformer models from scratch for text classification and NER, and a decoder-only recommender that suggests the next step while authoring tests.
  • PyTorch
  • Transformers
  • Computer vision
  • RAG
  • AWS
  • Kafka
  • Next.js
  • SvelteKit

Jan 2020 — Nov 2021

Machine Learning Consultant

Quinnox · Bengaluru

Built the NLP stack for automated content categorisation and a conversational bot for test automation.

  • Shipped text classification models with scikit-learn, FastText and Gensim, and transformer-based NER.
  • Designed a text augmentation process that injects realistic errors to harden models against noisy input.
  • Built a unified async backend (FastAPI, Flask, aiohttp) serving predictions from a fleet of models.
  • Python
  • scikit-learn
  • FastText
  • Transformers
  • FastAPI

Sep 2019 — Jan 2020

Data Science Intern

KrishiHub

Crop disease detection that went to production.

  • Wrote OpenCV algorithms to isolate diseased regions of crops and a meta-learning classifier for disease type, then integrated both into a deployed pipeline.
  • OpenCV
  • Meta-learning
  • Python

May 2019 — Jul 2019

Machine Learning Intern

IIIT Allahabad

Network traffic analysis across switches to optimise cluster efficiency, cutting cost and energy use.

  • Python
  • Data analysis

Jun 2015 — May 2019

B.Tech, Computer Science and Engineering

Vellore Institute of Technology

GPA 8.48/10. VIT Best Project Award, 2019. Published four papers as an undergraduate.

04Writing

Notes from the field

Long-form write-ups on agent architectures, MCP and building with language models.

Read on Medium

AI in Plain English

Schema-First Function Calling

Design the schema before the prompt: a more reliable way to get structured tool calls out of any model.

  • agents
  • function calling

Data Science Collective

MCP vs A2A: Building Bridges Between AI Agents

Where Anthropic's Model Context Protocol and Google's Agent2Agent overlap, where they do not, and when to reach for each.

  • mcp
  • a2a
  • agents

Data Science Collective

AI Engineer's Handbook to MCP Architecture

A multi-part walk through hosts, clients, servers, transports and the lifecycle of a tool call, written for people who ship.

  • Series
  • mcp
  • architecture

Generative AI

I Built an OpenAI-Style Swarm That Runs Entirely on My Laptop

Hand-off based multi-agent orchestration on Qwen 2.5, no API keys required. The write-up behind swarmloka.

  • agents
  • local llms

Towards AI

How I Developed a NotebookLM Clone

From PDF to a two-host podcast with GPT-4 and ElevenLabs. The story behind PDF2Pod.

  • audio
  • llms

Medium

Why you don't need LangChain for building a RAG bot

A plain-Python RAG pipeline, and why fewer abstractions made it easier to debug and ship.

  • rag

05About

Who is behind this

Bengaluru, IndiaHover or tap to develop

Seven years in machine learning, most of them close to the models themselves: I have trained transformers from scratch for classification, entity extraction and recommendation, fine-tuned detection and vision-language models, and quantized them to run on small hardware. Since 2023 the work has been mostly large language models and the agent systems built on top of them, but I keep coming back to training, most recently a compact vision-language model that reads a screen and decides where to act.

At QyrusAI I set the technical direction for AI. The through-line is reliability: a shared SDK so every agent runs the same loop with the same telemetry, a runtime that streams a session live and keeps the evidence afterwards, memory for long tasks, and evaluation that looks at the whole trajectory rather than just the final answer. I mentor the engineers building on it and plan delivery with product.

Outside work I spend a lot of time on agentic development itself: how far can a coding agent take a real product? Tinker, a bring-your-own-key coding assistant for VS Code, is the main one. The native macOS apps in my GitHub are experiments from the same question, built mostly to see what agents can ship.

Toolbox

Agents & LLMs

  • Multi-agent orchestration
  • MCP
  • Function calling
  • RAG
  • Fine-tuning (Mistral, Llama, Qwen)
  • Evaluation & LLM-as-judge
  • Prompt & schema design

Deep learning

  • PyTorch
  • Transformers
  • JAX
  • Computer vision (DETR, YOLO, Detectron2)
  • NLP & NER
  • Recommendation systems
  • Quantization

Languages

  • Python
  • Swift
  • TypeScript / JavaScript
  • SQL
  • Shell
  • Rust (learning)

Platforms & infra

  • AWS (Bedrock, SageMaker, Lambda, SQS, ECS)
  • Azure OpenAI
  • Docker
  • Kafka
  • Redis
  • DynamoDB
  • Playwright

Product & web

  • FastAPI
  • Next.js
  • SvelteKit
  • SwiftUI
  • Electron
  • Tailwind CSS

Publications

Google Scholar
  1. 01

    Malware Classification using Machine Learning Algorithms

    2nd International Conference on Trends in Electronics and Informatics (ICOEI)

    2018 · 32 cit.

  2. 02

    Big Data Technology in Healthcare: A Survey

    10th IFIP International Conference on New Technologies, Mobility and Security (NTMS)

    2019 · 4 cit.

  3. 03

    Classifying and Predicting DoS and DDoS Attacks on Cloud Services

    2nd International Conference on Trends in Electronics and Informatics (ICOEI)

    2018 · 5 cit.

  4. 04

    IoT based Smart Safe System

    International Conference on Computational Techniques, Electronics and Mechanical Systems (CTEMS)

    2018 · 2 cit.

06Contact

Let's build something useful

Open to interesting conversations, talks and collaborations

saglanivatsal@gmail.com

Or email me directly.