AVAILABLE FOR COLLABORATION 2026

PAVAN GANESH INDUPUDI // INFERNIX

Engineering
intelligent systems.

I'm Pavan Ganesh Indupudi (Pavan Indupudi) — known online as Infernix. I build AI systems where privacy, retrieval, and thoughtful software meet: privacy-preserving browser agents, multilingual RAG, Android apps, and developer tools.

Explore projects GitHub profile
FOCUSAI agents · RAG · Android
BASEAndhra Pradesh, India
01

About

Curiosity is the interface.
Building is the answer.

EDUCATIONB.Tech CSE, Andhra University
IDENTITYPavan Ganesh Indupudi · Infernix
FOCUSAI/ML · privacy-first agents · full-stack
BASEAndhra Pradesh, India

I move from a research question to a working prototype: testing assumptions, measuring results, and refining the experience until the technology is useful to real people. Recent work spans a privacy-preserving browser agent, a multilingual retrieval system with refusal guardrails, a media-investigation platform, and an on-device Android wellbeing app.

I care about systems that are honest about their limits — models that refuse when evidence is weak, privacy boundaries you can watch, and software that runs on your device instead of your data.

infernix@dev:~$ builder of things that matter.
02

Selected work

Four engineered systems with real benchmarks, real tests, and real repositories. Every number below is reproducible from the source.

PROJECT / 001 · FEATURED PUBLIC REPO · SIH 26171

Privacy-preserving browser agent

PRIVYSE

Your screen never leaves the device — and you can watch the gate.

Most computer-use agents ship raw screenshots — every email, ID, and password — to a remote vision model. PRIVYSE changes the order of operations: the page is understood and sanitized on-device, a fail-closed gate proves the payload is clean, and only then does the model reason.

PII F1 0.984 30/30 adversarial blocked 27/27 guard tests 0 MB raw data off-device

What I built: the tiered sanitizer and zero-leak gate, on-device MobileViT/BlazeFace/OCR perception, and the FastAPI + Ollama reasoning server with perception-driven model routing.

TypeScript · WXT · MV3FastAPIOllama · Qwen2.5-VLMobileViT · ONNX
View repository
PRIVYSE gate output: a KYC form where every sensitive value is replaced by a stable token before it reaches the model
Exact bytes that cross the privacy boundary
PROJECT / 002 PUBLIC REPO · HH GOA

Multilingual voice RAG

VOICE-RAG

Ask out loud. Get grounded answers — or an honest refusal.

A voice-first retrieval system across English, Hindi, Gujarati, and Telugu. Calibrated relevance floors refuse off-corpus questions instead of hallucinating; a grounding guard blocks any answer not traceable to retrieved text.

P50 56 ms 178 tests 4 languages

What I built: FastAPI pipeline with E5 embeddings over the 40K-passage MSMARCO-XI corpus, Sarvam speech-to-text, per-language refusal floors, and a zero-build vanilla-JS interface.

Python · FastAPIRAGSarvam STT
VOICE-RAG interface: a spoken question with evidence drawer, guard decision chain, and per-stage latency breakdown
Evidence drawer + latency trace
View repository
PROJECT / 003 PUBLIC REPO · UNESCO

Media & information literacy

SourceTrail

Don't just check the claim. Trace the story.

Built for the UNESCO "Play Your Part" hackathon. A deterministic investigation engine that never issues a True/False verdict — it reconstructs where media came from, what changed, and what evidence exists.

230 tests 11 golden cases NO_VERDICT engine

What I built: the A→G investigation engine — claim extraction, source discovery, reverse-image investigation, and evidence organization — plus the FastAPI backend and test suite.

Python · FastAPINext.js · TypeScriptVerification
SourceTrail investigation workspace with trace timeline, context comparison, and typed evidence panels
One-screen investigation workspace
View repository
PROJECT / 004 PUBLIC REPO · MIT

On-device digital wellbeing

Viora

Escalates gently — never bans.

An Android companion that watches usage rhythm and steps in with a calm nudge before a scroll becomes an hour. Zero network permission — no analytics, no servers, everything on-device.

What I built: MVVM + Clean architecture in Kotlin with Hilt, Room (5 tables, 90-day retention), a foreground usage-tracking service, Compose overlays, and the focus-session engine.

Kotlin · ComposeHilt · Room126 Kotlin files
View repository
Viora's four-stage intervention ladder: gentle nudge at 30 minutes, reflect at 60, recover at 120, protect at 180
The four-stage intervention ladder

Metrics are reproduced from each repository's own benchmark suites. Visit the linked source for methodology and current status.

03

Recognition

SIH 2026 · Track 26171
PRIVYSE
On-device visual perception for lightweight browser agents
UNESCO MIL Hackathon
SourceTrail
Youth media-literacy investigation platform
HH Goa 2026
VOICE-RAG
Multilingual RAG with refusal guardrails
HackAP · 2nd place
AgriSathi
Team project · 100+ teams · tribal advisory ecosystem
04

Engineering stack

LANGUAGES
PythonKotlinTypeScriptJavaScriptJavaC++
AI / ML & LLM
RAGPyTorchONNX RuntimeTensorFlow LiteVLMs (Qwen2.5-VL)Embeddings (E5)
FRONTEND & FULL-STACK
ReactNext.jsNode.jsExpressVanilla JS
BACKEND & APIS
FastAPIPydanticUvicornREST design
ANDROID
Jetpack ComposeHiltRoomDataStoreWorkManager
TOOLS & PRACTICES
pytestDockerBenchmarkingPrivacy-first designOpen source
05 / NEXT

Have a good problem?
Let's build.

Open to thoughtful collaborations, open-source projects, and conversations about AI and software engineering.

Connect on GitHub LinkedIn