AI engineer · full-stack & mobile · Karnataka, India

I build things that actually run.

I take ideas from zero to a deployed product: multilingual AI used in production, agents that ask a person before they act, and benchmarks that show exactly where models fail.

# dev-teamGatekeeper, live
Coding agent

About to run npm install rupee-words-formatter. Waiting for a verdict.

Scout

Stop: that package does not exist. The model invented the name.

Squat Watch

On the watchlist. If anyone registers it, I alert this room.

Fixer

Nothing safe to install. Writing the helper with no new dependency.

Coding agent

Understood. Not installing it.

69merged pull requests in Kestra
77AI tools in VerbX One, 76 languages
9build write-ups on DEV
12packages published on npm
01 / SELECTED WORK

Products in production,
not prototypes.

Each one is live, used, or open for you to try. Hover a card, then follow the links.

Co-founder & founding engineer · Ivish AI

VerbX One

One account, one wallet, 77 language tools in 76 languages: translation of text, documents, images, audio and video, app localization, voice agents on the phone, WhatsApp, meetings and a developer API. I built the inference backbone: FastAPI and ONNX pipelines for speech-to-text, text-to-speech, OCR, NER and emotion detection, with real-time cross-language calls over WebSockets and Redis.

77 tools76 languagesONNX inference12 npm SDK packagesApp on Google Play
NEW Flip the Switch · Nov 2026

Gatekeeper

Four agents in your team chat that check what coding agents install, before they install it. Invented packages, lookalikes, MCP servers, hidden instructions in rules files, curl | sh and pasted keys. Built on Switch.

84% malware caught (PyPI)Only a person can approveSlack · Discord · phone
Built solo, end to end

Prayana

An AI travel platform for Karnataka: day-by-day itineraries, offline mode, payments and an admin workflow. A full consumer product, not an itinerary generator.

79 screens42,000+ lines6-model fallback chainPlay internal beta
In build

Vigyl

A developer agent that lives on your server, notices a failed build, proposes the fix and waits for you. Nothing is written until a person sees the diff and approves it, from the web or from Claude over MCP.

@vigylx/cli on npmHuman approval gate
02 / RESEARCH & EVALUATION

Where models fail,
measured.

A prompt set, a live check against ground truth, hand spot-checks, and a public write-up. Tap a card.

GHOST PACKAGES · GEMMA 4

An invented package, asked for again

3 of 4

150 realistic coding prompts to gemma-4-31b-it, every install name checked live on PyPI and npm. One invented name came back in 3 of 4 re-asks: predictable enough for an attacker to register. Ships ghostcheck, a guard that flags ghost dependencies before install.

KAGGLE BENCHMARK · 11 MODELS

Indian GST, UPI and lakh-crore

11

A public Kaggle benchmark on the formats most evaluations skip: GST rules, UPI references and lakh/crore numbers. Several famous models got Puducherry's GST details wrong.

DAWAIGRAPH · DRUG INTERACTIONS

Grounding answers in a graph

40%

160K interaction pairs and 70K Indian medicine brands. On 30 real regimens, an LLM alone found 48 of 119 known interactions and made 52 claims the reference does not list; graph-backed answers stayed faithful to the source.

GATEKEEPER · HELD-OUT MALWARE

Reading packages without running them

1 in 500

On real malware it had never seen, Inspector flags 84% on PyPI and 60.5% on npm, and wrongly flags only 1 in 500 popular npm packages. No AI model makes the call: live registry checks and code read as text.

03 / WRITING

Every build,
written up.

Nine long reads on DEV: what I built, the numbers, and what broke on the way. Newest first.

04 / OPEN SOURCE & BUILDS

Shipped in public,
on deadline.

My own workspace.Six small apps with one shared backend that I use to run my own days. Prompted on Lovable, on purpose: speed for myself, not a product.

Ping ↗ Inbox ↗ Memory ↗ Search · in buildBrowser agent · in buildAI gateway
05 / EXPERIENCE

Where I've worked.

Co-founder & Founding Engineer

IVISH AI QUANTUM PVT LTD · DEC 2024 – NOW

I work across the company's AI and product ecosystem, with most of my time on VerbX: AI and ML inference, translation, speech-to-text, text-to-speech, OCR, NER and emotion detection, backend APIs, mobile apps and cloud infrastructure.

I built backend services in Python and FastAPI, ONNX inference pipelines, and real-time cross-language communication over WebSockets and Redis Pub/Sub. Everything now ships as VerbX One.

06 / TOOLKIT

What it's built with.

B.Tech, Smart Manufacturing, PDPM IIITDM Jabalpur.

Languages
TypeScriptPythonJavaScriptC++SQL
AI & ML
LLM evaluationAI agentsMCPONNXPyTorchSTT · TTSOCR · NERTranslation
Web & mobile
ReactNext.jsReact NativeExpoTailwindRemotion
Backend & cloud
FastAPINode.jsWebSocketsRedisPostgreSQLMongoDBDockerGoogle CloudAWSFirebase

Let's build
something real.

Hiring, collaborating, or want a second pair of eyes on an agent or an eval? I reply to every message.