Hi, I’m Ayush, an AI engineer.

I build AI agents and the systems they run on.

Agents that use real web and mobile apps and read whole codebases, and what sits under them: the model layer, the evals that keep them honest, and deployments on three clouds.

I started with neural networks in college: face recognition, image classifiers, and a small drone that learned to follow a face.

See what I build

A bit about me

It started with a broken computer.

Fifth standard
I was deep into Call of Duty 4 when our PC died. It wouldn’t start, it just made a noise. I was scared of getting a shock, but I opened the cabinet anyway and slowly took everything out: the RAM, the hard disk, the cables, the fan, the CPU with all its tiny pins. I didn’t know what any of it was called yet. I cleaned the RAM and its slot, the CPU and the dusty fan, put it all back, and it started.
Before we had internet at home
A salesman came to the door with two books, and I talked my mom into them. “Maybe you’ll learn something,” she said. DK’s Ultimate Family Visual Dictionary taught me what a CPU, RAM and a hard disk were. The Knowledge Encyclopaedia covered science, medicine and space, and kept me asking questions.
College
I found software, then neural networks. Building things became how I learned: face recognition first, then a drone that follows faces, then a paper on a network that picks outfits.
Now
I build AI agents and the systems around them, still opening things up to see where they break. I overthink everything and have high standards. My friends would tell you that, and they’re right.

I care more about building things that work than collecting credentials for things that sound impressive.

What I build now

The agent is the easy part.

Making an agent useful means building everything around it, from the layer it thinks through to the platform it ships on.

Agents that use software
Agents that operate real web and mobile apps from the DOM, the screen and the native view tree, with vision-based clicking when a page won’t cooperate.
seeDevika Browser assistant
Agents that read code
Pipelines that read a repository, map its features and write test cases from the parsed source, and that work out which tests a pull request affects. The model has to cite what the parser found, so it can’t invent a route.
The model layer
One layer between every feature and every model: Gemini on two APIs, GPT on Azure and Bedrock, each task with its own model, thinking level, timeout and fallback. Every attempt’s tokens and cost are recorded, failed ones too.
Evals and observability
Benchmarks with known answers, blind audits, and replays of recorded sessions on another model before anything changes. I built the console that shows what each agent run did, how long it took and what it cost.
The platform under it
One codebase that runs as SaaS and on-prem on Google Cloud, AWS and Azure. A job queue with leases and crash recovery, graceful drains on deploy, keyless cloud identity, and a preview environment for every branch.
Where people already work
The MCP server people use from Claude, Cursor and VS Code, the GitHub check that says what a pull request affects, and a Slack assistant for my team.
Neural networks, where it started
CNN image classifiers, transfer learning on VGG16, face recognition, YOLO object detection on a drone, and a multi-output network for outfit recommendation that became a paper.
seeTello drone The paper

Experience

Four places, all building with AI. Newest first.

Full background on LinkedIn

CQ
ContextQA
An AI platform for software quality
1. Senior AI EngineerNow
  Oct 2024 - Now
  I wrote the first commit of the AI backend in May 2024 and still own it: I review the team’s changes, wrote its coding guidelines and run its releases.
  - Built the agents that run tests on real web apps: a planner and a decision loop with action, verify, extract and ask agents that work from the DOM and screenshots, with vision-based clicking as a fallback. Extended them to native Android and iOS apps.
  - Built code analysis. It parses a repository’s routes and UI with tree-sitter, sends an explorer agent through large codebases in a network-blocked Daytona sandbox, and writes feature-level test cases that are linted, repaired and independently reviewed. Missing facts are recorded as gaps, not invented.
  - Built change-impact analysis with a teammate. It maps a pull request to the existing tests it affects, tuned against a recall and precision eval.
  - Also built the model layer every feature runs through, merged the SaaS and on-prem codebases into one image for three clouds, and built the console the team uses to debug agent runs and LLM cost.
2. AI Software DeveloperPart-time
  Apr - Sep 2024
  - Built the first version of the browser extension that generates test steps and runs agent actions on the page.
  - Built a testing agent that turns plain requests like “test the login flow” into test runs, with DOM analysis, HTML shrinking, and action checks for complex B2B web apps, using vision models.
  - Built debugging tools with video logs, screenshots, traces, and automatic root-cause notes for failed runs.
- Python
- Flask
- Gemini
- Vertex AI
- OpenAI
- tree-sitter
- Daytona
- Redis
- MongoDB
- Qdrant
- Next.js
- GKE
- AWS
- Azure
MM
Mirai Minds
1. AI Engineer
  Feb - Oct 2024
  Built a browser assistant, company-knowledge search, and cloud scraping services. Worked on backend latency and reliability.
  - Built a browser companion: a multi-agent system inside a Chrome extension that decides and acts on web pages.
  - Built a retrieval engine that indexes PDFs, Word files, websites, and Twitter data into Pinecone and FAISS for a customer chatbot.
  - Built a scraping system on Playwright that runs many browser instances at once on a cloud server.
  - Worked with GPT, Claude, Phi, LLaMA, and Mistral, choosing the model for each task and tuning the backend for low latency.
- LangChain
- Pinecone
- FAISS
- Playwright
- Chrome extensions
WW
Whatiwear
My startup, never launched
1. Co-founder & CTO
  2023 - Feb 2024
  A fashion and AI startup I worked on in my last year of college and until February 2024. We never launched it.
  - Built the backend, web app and admin panel, and collected feedback from real users along the way.
  - Trained the models behind it: a CNN clothing classifier, VGG16 transfer learning, and the MobileNet style classifier it shipped with, plus body-shape and skin-tone analysis with MediaPipe.
  - Met fashion designers to learn how fashion actually works.
- Python
- TypeScript
BD
Blue Data Consulting
1. Cloud Consultant / MLOps Intern
  Jan - Apr 2023
  Trained deep learning models on SageMaker and Azure ML, built chatbots, and deployed ML services on AWS.
  - Built banking and IT-support chatbots on GPT-2 Medium and BERT with Hugging Face, and trained VGG16 and VGG19 image classifiers.
  - Used YOLOv5 to find signatures and photos on documents, and Tesseract OCR to pull data out of invoices and forms.
  - Built a job recommender on word2vec. Its training text came from a parallel request system (asyncio and aiohttp) that sent about 180 requests a second to OpenAI’s API.
  - Studied the cost of serving models on EC2 versus Lambda, and deployed a containerised Flask app on ECS Fargate.
- AWS SageMaker
- EC2
- Lambda
- Azure ML
- Hugging Face
- YOLOv5

Projects

Open source I helped run, and things I build on my own time, from a drawing assistant to a drone.

Devika

Open source · former maintainer, 2024

Repository

An open-source AI software engineer, built as an alternative to Cognition’s Devin.

You give it a high-level instruction. It breaks the work into steps, researches what it needs on the web, and writes the code. I was one of its maintainers in 2024, and its second-largest contributor by commits.

What I worked on

Added Gemini and Mistral as model providers, and fixed the Ollama integration for local models.
Fixed live progress updates in the interface, along with a round of UI and backend improvements.
Made model output sturdier: a better response parser, handling for invalid responses, and timeout controls.
Kept API keys out of the logs, and rewrote the installation guides.

Python
Flask
Socket.IO
SvelteKit
Playwright
Claude
GPT-4
Gemini
Mistral
Ollama

GitHub stars: 19.5k stars
Forks: 2.5k forks

Devika writing a seven-step plan to add Pinecone embeddings with LangChain — Devika screenshots

Devika browsing the Pinecone Python docs while it works through its plan — Devika screenshots

Tello drone

Computer vision · built in college, 2022

A DJI Tello that flies itself.

The drone streams video to my laptop, which does the heavy work: finding faces and objects in each frame, then sending flight commands back.

Face detection and tracking that keeps a person in frame, plus YOLOv3 object detection and edge detection.
Voice control in English and Gujarati. Say “ઉપર જા” and it climbs. It answers with pre-recorded Amazon Polly audio.
A keyboard flight mode, and live recording from the drone’s camera.

In progressRebuilding it around a language model that sees the video and takes spoken requests.

Python
djitellopy
OpenCV
face_recognition
YOLOv3
SpeechRecognition
Amazon Polly

Source Write-up Video

Excalidraw AI

Personal project · 2025

An AI assistant I built into a fork of Excalidraw.

Describe a diagram, like “draw a user authentication flow”, and it appears on the canvas, fully editable. Paste a screenshot of any diagram and it rebuilds it as a new Excalidraw drawing.

It can also change existing elements and answer questions about Excalidraw. The editor itself is the Excalidraw team’s work; I built the assistant.

TypeScript
React
LLMs
Vision models

Source Demo

Browser assistant

Built at Mirai Minds · 2024

A Chrome extension that does web tasks for you.

Filling in forms, gathering information, and repeating tedious steps, all from a single prompt.

I worked on its multi-agent assistant and backend services, along with retrieval over company documents and websites, a cloud scraping system, and backend latency.

There was also a workflow version: every step the agent took became a node on a canvas, with a screenshot, and the whole flow could be played again.

Chrome extension
Multi-agent
LangChain
Playwright

Demo

Whatiwear

Startup · 2023 - Feb 2024 · rebuilt in 2025

An AI stylist I co-founded. We built it and tested it with real users, but never launched it. It follows one thread: my undergraduate paper on outfit recommendation (May 2023), then a virtual try-on experiment, then Whatiwear.

From the 2025 rebuild: the AI tags each piece of clothing you upload.

Earlier experimentsVirtual try-on CP-VTON+ on PyTorch 2 My first face-recognition project

How I build AI systems

The rules I keep coming back to, mostly learned from agents that got something wrong on a real app.

Ground the model in evidence
Agents decide from what is really there: the DOM, screenshots, parsed routes, retrieved tests. Not from a guess about the page.
MeasuredMade a model choose from what a parser had already found instead of writing freely: invented answers 52% → 0%, 9.6× fewer tokens, and one step went from 22 min → 3 min.
Record the gap, don’t invent
If a fact isn’t in the source, the pipeline says so instead of writing a test for it.
MeasuredOn one open-source app, 36% of generated test cases checked for translation keys as if they were visible text. A lint now holds those back instead of shipping them.
Fix the general case
When a model gets something wrong, I fix what it misunderstood, not the one page it failed on. A patch for one app is a bug for the next.
MeasuredOne shared sign-in step instead of one per test: “missing prerequisite” failures 92–100% → 0%.
Measure on real runs
Benchmarks with known answers and replays of recorded sessions decide changes, not one good run. Every call’s tokens, latency and cost are tracked.
MeasuredOn a fixed benchmark of 58 scenarios, coverage went 7 → 38–41, with run-to-run noise measured at ±3 before trusting any gain.
Never overstate the result
A partial run says it’s partial. Counts show what was delivered, and a zero says whether anything was actually searched.
MeasuredFound runs reporting about a third of their real cost, because some loop steps and thinking tokens weren’t counted. Fixed, then checked against production: zero difference.
Plan for failure and load
Token and time budgets, bounded retries, partial results kept, a fallback when something new breaks, and no single request allowed to take a server down.
MeasuredAn endpoint that was running servers out of memory: 37 s → 1.6 s, and 150 MB → 81 KB per request.

Measured lines come from my own work in 2025–26: recorded runs, fixed benchmarks and production data. I’ve left the products and features unnamed.

AI workflow

You don’t have to do it by hand anymore.

AI can already do a lot of the work most of us still do manually, and I want more people to know it. Here’s a small, everyday example: ChatGPT worked out what to update on my LinkedIn profile, and Codex, driving my Chrome, made the edits. I didn’t write or change anything myself. I only reviewed.

My own work runs the same way: several Claude Code and Codex sessions side by side, each on its own branch, and me reviewing what comes back. This site was built like that too.

Read the full post

Slack + GitHub
Gather the context
ChatGPT
Draft the changes
Me
Review them myself
Codex + Chrome
Apply in the browser
LinkedIn
Check what saved

Open source & community

Work outside my own repositories, and time spent with other builders.

Open-source contributions

Community & writing

Education & research

B.Tech, Computer Engineering

P. P. Savani University, Surat

Focused on machine learning and computer vision, which led to the research paper here.

Years: 2019 - 2023
CGPA: 8.97 / 10

Undergraduate research

Enhancing Fashion Recommendations: Deep Neural Networks for Personalized Outfit Suggestions

With Abhijitsinh Parmar · IJSRST, Volume 10, Issue 3

We proposed an AI stylist model built on fundamental fashion theory and earlier work on AI in fashion, and looked at how neural networks are implemented and perform in the fashion domain.

The model is a small multi-output network: body measurements and skin tone go in, one shared layer learns from them, and five heads score T-shirt colour bands, trained on outfits labelled with fashion designers.

The same interest in fashion and AI carried into Whatiwear, the startup I co-founded.

May 2023

DOI 10.32628/IJSRST523103117

Read the paper