Bhanu Nagpure — Software Engineer & Indie Founder

01 · apps

Products under Devanshu Studios.

Two production releases. One in Google Play closed testing. Each one: architecture, monetization, CI/CD, Crashlytics, and publishing — owned end to end.

Keepary

com.devanshustudios.keepary · Flutter · v1.0.3

closed testing

Scan to PDF on the phone. OCR, merge, compress. No account. On-device OCR with Google ML Kit — zero API cost, zero cloud. Local document vault with on-device search, PDF tools (split, merge, compress to target size, password), ID-card mode, QR scan, and signing. Keepary Pro via RevenueCat: searchable PDFs, encryption, high compression. Ads engineered with caps and cooldowns — interstitials only after saves, never on first launch.

FlutterRiverpod 3ML Kit OCRRevenueCatAdMobCameraoffline-first

Studio on Google Play ↗

PyMaster

com.devanshustudios.pymaster · Flutter

live on play

Learn Python on the phone. 1,200+ exercises across gamified levels with an AI tutor, offline-first sync, and a clean-architecture Flutter codebase past 50k lines — built solo. RevenueCat subscriptions + AdMob native ads in production; the monetization postmortem is on the blog.

FlutterRiverpod 2.0FlavorsFirebaseRevenueCatGemini tutor

Get it on Google Play ↗

Chakuli

com.devanshustudios.chakuli · Kotlin · Jetpack Compose

live on play

Privacy-first on-device AI chat — with a dual runtime. Runs modern LLMs entirely on the phone: Gemma 4, Qwen3.5, Phi 3.5, and SmolLM — switchable between LiteRT and GGUF engines with per-model accelerator profiles and dual sampler modes for thinking vs. instruct. No cloud inference, no accounts. Jetpack Compose UI, Hilt, CameraX vision skills, and agent skills with per-call permissions.

KotlinComposeLiteRT-LMGGUFGemma 4 · Qwen3.5HiltCameraX

Get it on Google Play ↗

02 · hackathons & case studies

I don't just attend hackathons — I measure what I build.

Four hackathons this season: WeMakeDevs × Modiqo, the Zerops Challenge, Bright Data's Into the Scrape-Verse, and the Agent Harness Hackathon. One of them turned into a full research study.

Rote Playoffs — the measured one

WeMakeDevs × Modiqo · Sep 1–6 2026 · 38 Plays published, all scored 1.00 · top-5 of 123 makers

I shipped a 38-Play arsenal of zero-key agent procedures, then asked the question every builder should ask: does crystallized procedure memory actually beat model effort? The answer became a controlled study — same agent, same model, same task, with-Play versus re-derivation. Frozen ground truth, deterministic grader, six tasks, thirty-one graded trials, a leak audit that caught two baselines reading the experiment's own answer key (reported, not hidden), and raw telemetry for everything.

9.9×

faster with a Play

15.3×

fewer output tokens

9.8/10

correctness, replicated 6 tasks

Open the full study → Artifacts on GitHub ↗

transfer task	w/ Play	baseline
side-effect audit	10, 10	7, 7
instruction burial	9, 9	2, 2
commit hygiene	10, 10	8, 8
JSON integrity	10, 10	10, 9
event deadline join	10, 10	10, 10

baseline @ 572–2,180 s vs with-Play @ ~32 s on instruction burial

Crucible

The Zerops Challenge · WeMakeDevs × Zerops

Your URL under fire, live: four challenge packs (heartbeat, HTTPS/uptime, JSON validation, capped concurrent load) with a real scoreboard and letter grades. Fastify + React 19 + PostgreSQL with an async probe worker as the product — deployed on Zerops, SSRF-hardened, pre-seeded scorecards so it's never empty for judges.

github.com/bhanu-dev82/Crucible ↗

HYDRA — shrinkflation watch

Into the Scrape-Verse · Bright Data

"The jar got smaller. The price didn't." Custom Scraper Studio collectors read ALDI, Sprouts, and a self-hosted fixture shelf; when package size drops and price holds, HYDRA flags it — and collectors self-heal from a plain-English field list when the HTML moves. Live on Vercel with a 1:38 demo video.

Live demo ↗ · repo ↗

Agent Harness

The Agent Harness Hackathon

A quota-aware multi-tier Gemini model router, an agent-harness client with turn running, local + remote MCP tool configs, and automated Qodo PR-review workflows — packaged with offline docs, project starters, and a Kaggle benchmark harness so any agent (or human) can execute the build cold.

github.com/bhanu-dev82/agentscope ↗

03 · experience

Where the work happened.

Founder & Lead Engineer

Devanshu Studios · Remote · Aug 2025 – Present

Shipped PyMaster (50k+ LOC, live) — Clean Architecture, Riverpod, Flavors, Firebase, RevenueCat, Gemini AI tutor. Shipped Chakuli — on-device Gemma/Phi/SmolLM via LiteRT-LM, Compose, Hilt, CameraX. Leading Keepary through closed testing: architecture, monetization (AdMob + IAP), CI/CD, Crashlytics, Play Console publishing.

Open Source Contributor

Wikimedia Foundation — Scribe-Android · Remote · Feb 2025 – Present

Tablet keyboard layouts in Kotlin for 7+ languages; UI fixes and a KeyHandler refactor for maintainability.

Machine Learning Intern

Athabasca University · Canada · Aug – Oct 2024

95%-accuracy ideology classifier (BERT/Transformers) plus a researcher web UI for validation.

Globalink Research Intern

Mitacs · Edmonton, AB, Canada · Apr – Aug 2024

Sole developer on a cross-platform Flutter field-trip app (iOS/Android/Web) with Firebase and OpenStreetMaps; presented at the TRESL Lab Symposium.

Software Developer Intern

BETIC-GHRCE · Nagpur · Oct 2022 – May 2024

TMJ screening app with MediaPipe FaceMesh + ML Kit — 60% accuracy gain, 30% camera speedup. Contributed to a granted patent (Pub. No. 202321022476).

Fellowship — Design for Health

IIT Bombay · Mar – Oct 2022

Led a medical-device prototype for xerostomia — 80% faster assessments; commended at IIT Bombay.

04 · projects

Research, engineering, impact.

V-Safe-Anywhere

Women's safety wearable + Flutter app: panic button triggers live GPS sharing via Firebase. 98% uptime. Published by IGI Global.

Flutter · Firebase · GCP · IoT — IGI Global 2024

OCanPredict — OSMF Screening

Android app using MediaPipe FaceMesh for OSMF diagnosis — 40% efficiency gain. National awards + patent contribution.

Kotlin · MediaPipe · Firebase — Pub. No. 202321022476

Autonomous Garbage Collection Robot

YOLOv5-powered robot for floating waste detection and retrieval — 60% less manual effort. ACM ICACS 2024, Hong Kong.

Python · YOLOv5 · OpenCV — ACM 2024

05 · stack

What I ship with.

Mobile is a pillar — not the whole house. Production stack across product, systems, and ML.

mobile

KotlinFlutterJetpack ComposeDartRiverpod 2.0MVVM / MVIHiltCameraXWorkManagerRoomPlatform Channels

ai / ml

LiteRT-LMGemma 4Phi 3.5SmolLMMediaPipeML KitBERT / TransformersTensorFlowYOLOv5Gemini API

systems & backend

Firebaseoffline-first syncRESTHiveProto DataStoreSQLGoogle CloudLinux

product & ship

Clean ArchitectureRevenueCatAdMobPlay ConsoleFlavorsCrashlyticsGit / GitHub

06 · recognition

Proof the work left the lab.

Awards & grants

Third Place — Avishkar 2024 · Runner-up — DIPEX 2023 · Runner-up — SHRISTI 2023 · Grant — BIREC E-YUVA (₹10 lakh) · Winner — MEDHA & MEDIC-2022

Publications & patents

"V-Safe-Anywhere: Wearable AI & IoT" — IGI Global, 2024. "Autonomous Floating Garbage Collection" — ACM ICACS 2024. Patent (granted): Smart Chairside TMJ Examination Tool. Patent (pending): AI Portable Device for Dental Disease Detection.

Conferences

"Mobile Field Trip Application" — TRESL Lab Symposium, Athabasca, Canada, 2024. "Autonomous Garbage Collection" — 8th ICACS, Hong Kong, 2024.

I ship end-to-end products —
mobile, on-device AI,
and the systems under them.