Self-driven engineer who learns by building. I research a problem, then ship the fix: on-device AI apps for Android, LLM and agent workflows, and enterprise automation, on a mechatronics foundation (8.29 CGPA). Based in Komarapalayam, Tamil Nadu, India.
AI-focused engineer with five released Android apps, a live web app and an open-source AI agent skill:LiveSub runs speech recognition and translation fully on the phone, Offline Reel Captions makes offline subtitles for Tamil, Tanglish and other Indian languages, Artisan NPR Studio turns photos into art with 134 GPU styles, Live Wallpaper Maker lets anyone design layered live wallpapers with tilt parallax and 3D transitions, Net One Tap resets the mobile network in one tap, ASIAN MOBA 5v5 Tournament runs sign-ups, teams and brackets for a Mobile Legends tournament on Firebase, with Gemini reading players' stats screenshots, and Preflight Agent Skill makes AI coding agents research before they build and learn a lesson from every bug. I work with LLMs every day, from API calls with structured prompts to file-based agent workflows (ICM), backed by enterprise NetSuite automation experience (SuiteScript 2.1, Coupa, Filevine, Litify) and a hands-on robotics and embedded systems background. Most of what I use today I taught myself, from docs, source code and research papers.
Module 01 — Apps, AI & Hardware
Live subtitles for anything your phone plays.
Edge AI • Sherpa-ONNX • ML Kit • Jetpack Compose
A private, offline speech-to-subtitle translator for Android. It listens to a video, stream, game or call, turns speech into text, translates it, and shows the result in a floating subtitle window over any app. Speech recognition and translation both run on the phone: no account, no API key, no audio uploaded.
AudioPlaybackCapture (or the mic), resampled to 16 kHz with an anti-aliasing filter and a voice clean-up stage tuned for speech.
Mobile internet stuck? Fix it in one tap.
Android • Kotlin • Shizuku / Root • Telephony
Restarts only the mobile radio, the same fix as toggling airplane mode, while Wi-Fi, Bluetooth and mobile data stay untouched. Also shows a per-SIM 5G status notification.
ITelephony.setRadioForSubscriber through Shizuku or root, with a soft-refresh fallback. Skips the reset during calls and always brings the radio back.
ServiceState (the same values as the *#*#4636#*#* screen) with dual-SIM support and no location permission.
Auto subtitles for Tamil, Tanglish and more, made on the phone.
Whisper • IndicTrans2 • ONNX Runtime • Media3
Pick a video and a language, and the app writes the captions: in the original script, in English letters by sound ("vanakkam nanbargale"), or translated. Fix and style every caption, then export an MP4 with the captions burned in, or an .srt file. Built for South Indian creators. The video never leaves the phone.
Turn any photo into a sketch, painting or comic.
Kotlin • Compose • OpenGL ES 3.0 • GLSL
134 art styles in 12 groups: pencil sketch, oil paint, watercolour, halftone, pixel art, film and glitch looks. Each style has its own controls, and a high-resolution export looks just like the preview. The APK is about 2 MB and asks for no permissions.
Design your own live wallpaper: layers, tilt parallax and 3D scene transitions.
Kotlin • Compose • OpenGL ES 3.0 • MediaCodec
Stack videos, GIFs, photos, gradients, particles and a live clock, give each layer its own tilt and swipe depth, add looks and keyframe animation, and switch between scenes on a timer, on unlock, on double-tap or by time of day, with crossfade, 3D flip and cube transitions. Version 1.1 adds one-tap 3D photos (on-phone subject cut-out with ML Kit, the clock tucked behind the subject), water-ripple tap effects, separate lock-screen scenes, motion photos, animated stickers, and sharing as video or GIF.
Sign-ups, teams and brackets for a Mobile Legends tournament, in one phone-friendly web app.
React • TypeScript • Firebase • Gemini Vision
Players sign in with an email link, confirm their game ID, upload three stats screenshots and pay the entry fee by UPI. They form 5-player teams with a 6-digit code, then play through a seeded bracket, shown as a tree with colour-coded best-of rounds, and report each result with screenshots. On match day each team's page shows the next opponent and start time, players check in, and winners get confetti until their next match. Admins handle payments, reviews, refunds, brackets, disputes and emergency substitutes from one dashboard, and every admin change is logged. The look is red, matte black and gold, with a shimmering gold lion logo.
Makes AI coding agents research before they build, and learn from every bug.
Agent Skills (SKILL.md) • Claude Code • Codex • Gemini CLI
A reusable skill for AI coding tools. Before the agent picks a library, API, SDK version or AI model, it checks the official docs for deprecation, device support and licence. After every snag, it saves one short "before doing X, check Y" lesson that it reads on the next project, so the same mistake doesn't happen twice. My own copy started with lessons from the snags in all my apps above.
session.md per project so any new session can pick up the work.
Personal Project • Termux • Obsidian
A personal AI assistant running on my phone in Termux, with long-term memory kept in an Obsidian Markdown vault. Notes and their links form a knowledge graph the assistant retrieves from (graph-based RAG), so context carries across sessions.
Mechatronics Capstone • 2023
A remote-controlled wall-climbing robot that cleans cobwebs off high walls. Dual high-RPM BLDC thrust fans create a low-pressure zone that holds it to vertical surfaces.
ESP32 • Python • IoT
Developed an IoT predictive maintenance system calculating dust accumulation rates on air filters. Transmitted telemetry to ThingSpeak, reducing scheduled servicing overhead by 25%.
Module 02 — Research & Self-Learning
Net One Tap started as a research write-up: what the toggle really resets, what Android allows an app to do, which hidden APIs work on which versions, and where the phone's 5G status values come from. The app came after the answers.
For LiveSub I compared speech models (SenseVoice, Moonshine, Parakeet, GigaAM, Whisper and streaming models) by language, speed, size and license, then matched each one to the job: streaming for instant text, whole-phrase models for accuracy.
I run my AI-agent work with ICM (Interpretable Context Methodology): numbered folders and Markdown files act as the pipeline, so every step's context is visible, editable and reusable, instead of hidden inside framework code.
Android, Kotlin, on-device AI and LLM tooling were not in my syllabus. I learned them from documentation, open-source code and papers, and proved it the only way that counts: released apps people can download.
Module 03 — Data Pipelines
Associate Developer / NetSuite Consultant • Jan 2026 – Jun 2026
Promoted from Technical Intern • Sep 2025 – Jan 2026
Mechatronics Engineering Intern | Hosur, India
Assisted with mechanical calibration, end-effector testing, and cell operations on FANUC industrial robots. Implemented C programming and PLC ladder logic for machine diagnostics.
Coding Coordinator & Campus Leader
Coordinated competitive coding challenges and infrastructure for campus symposiums Nexus (2022) and Throb (2023). Mentored 20+ peer students in foundational C programming, Python, and data structures.
Module 04 — Educational Metrics
Kongu Engineering College | 2020 - 2024
S.S.M. Lakshmi Ammal Matric Higher Sec School, Komarapalayam
S.S.M. Lakshmi Ammal Matric Higher Sec School, Komarapalayam
Module 05 — Skills & Honors
LLM APIs with structured prompts and outputs (Python), agentic workflows with ICM (Interpretable Context Methodology: folders and Markdown files as the agent pipeline), AI-assisted development with coding agents (Claude Code), reusable agent skills in the open SKILL.md format, Gemini vision on Vertex AI (reading screenshots into structured data), graph-based RAG over Markdown knowledge bases, Generative AI, Hugging Face.
Sherpa-ONNX speech recognition (streaming and whole-phrase), Whisper, Silero VAD, ONNX Runtime, IndicTrans2 and Google ML Kit translation, Kotlin, Jetpack Compose, Coroutines, Room, DataStore, Media3, MediaCodec / MediaMuxer video pipelines, OpenGL ES 3.0 / GLSL shaders, live wallpapers (WallpaperService), motion sensors, foreground services, overlays, Shizuku and hidden Android APIs.
React, TypeScript, Vite, Tailwind CSS, Firebase (Hosting, email-link Auth, Firestore, Cloud Storage, Cloud Functions), Firestore and Storage security rules, Cloud Scheduler, Google Cloud / Vertex AI, Firebase Emulator Suite, Zod, Vitest.
SuiteScript 2.1, SuiteAnalytics, SuiteFlow, Saved Searches, RESTlets, Suitelets, Coupa, Filevine, Zapier, OAuth 2.0 / TBA, Google Apps Script.
FANUC Robotics, Allen-Bradley PLCs, SolidWorks, ESP32, Arduino, LabVIEW, PCB Design.
Python, Kotlin, TypeScript, JavaScript, Java, C, Embedded C, SQL, Gradle, ADB, Linux / WSL2, Git, GitHub.
Module 06 — Focus & Discipline
Developing hyper-focused spatial awareness, sub-frame reactions, and discipline through punishing mechanical combat designs:
Currently open to new roles, while building new projects. Happy to talk about AI engineering, on-device AI, Android apps, agent workflows, and enterprise automation.
Open to new roles
Pick whatever is easiest for you.
IN-THRUST Robot • Prototype demo