class VarunAditya:
# ─── Identity ─────────────────────────────────
institution = "R.V. College of Engineering, Bengaluru"
program = "Information Science Engineering · Yr 2 / 4"
cgpa = 9.64
# ─── What I Build ─────────────────────────────
domains = [
"🧠 AI/ML Engineering & LLM Pipelines",
"⚙️ Full-Stack Systems Architecture",
"📱 Mobile & Edge Intelligence",
"🤖 Multi-Agent Autonomous Systems",
"🔗 Blockchain & Decentralised Apps",
"🖥️ Systems Programming & OS Dev",
]
# ─── Right Now ────────────────────────────────
active = {
"building" : "INDRA · NyayaSetu · Android AI Agent",
"researching" : "LLM hallucination mitigation, "
"context-aware architectures",
"mentoring" : "130+ engineers · Web Dev Bootcamp · RVCE",
}
# ─── Philosophy ───────────────────────────────
philosophy = "Build systems that reason. Not tools that respond."
def current_mission(self):
return "Architect AI that works at the intersection " \
"of intelligence, performance, and scale."
 INDRA · NyayaSetu · Android AI Agent |
 Mult-Agent Architectures |
 130+ Engineers · Coding Club RVCE |
 LLM Architectures · Edge AI · Multi-Agent |
Real-time power quality intelligence for substation operators — built for Hitachi India R&D.
The Problem: TXpert Hub collects transformer sensor data. Operators wait minutes for analysis, missing intervention windows measured in seconds.
What I Built:
- 🔬 Ensemble ML Core — 1D CNN learns latent waveform patterns; XGBoost classifies 31 handcrafted features. 60/40 weighted fusion identifies 8 fault classes in 20ms windows
- 🔩 Physics Validation Layer — pandapower digital twin cross-checks every ML prediction against IEEE C57.91 thermal models and IEC 60599 DGA ratios — preventing false alarms
- 🧠 Waveform Memory (FAISS) — 64-dim embeddings retrieve top-3 similar historical faults with verified root causes and past operator actions
- 💰 Real-Time Rupee Counter — transformer aging, reactive waste, and failure risk quantified live in ₹ so operators understand urgency without reading graphs
- 🏛️ Compliance Engine — live IEC 61000-4 + CEA scoring; auto-generates PDF compliance reports when score dips below threshold
- 🖥️ Premium Control-Room Dashboard — live waveform canvas, 3-zone incident feed, WebSocket telemetry, oscilloscope view
My Role: Team Lead & Full-Stack Developer — owned the complete technical stack from operator dashboard to ML inference backend.
Impact: End-to-end latency 200–400ms from waveform to operator recommendation. One prevented transformer failure (₹15–40 lakh) pays for deployment 50× over.
🥈 2nd Place · VISISONICS AI'26 · MIT Bangalore × Hitachi India R&D Centre
Turning India's 5.39 crore pending court cases into something understandable for people with a Grade 5 education.
The Problem: 80% of Indians are eligible for free legal aid. Almost none receive it. Legal documents are written for lawyers — not the people they affect.
What I Built:
- 📱 Android App (Kotlin + Jetpack Compose) — CameraX document capture, live clause risk alert cards, scheme entitlement cards, voice input for Kannada and Hindi
- 🎤 Voice-First Interaction — full speech input + audio output pipeline with multilingual TTS and STT, making the app usable without literacy
- 🔍 Document Intelligence Pipeline — Mistral OCR → Gemma 3 simplification → DistilBERT risk classification → IndicTrans2 translation; each section gets a readability score and retry if above Grade 6
- 📚 RAG Legal Q&A — ChromaDB vector store over Indian Kanoon precedents + NALSA schemes; users can ask any question by voice and get a grounded, cited answer
- 📋 Scheme Matcher — surfaces BOCW, PMJAY, PM Awas Yojana entitlements based on user profile attributes
- 📡 eCourts Live Tracking — direct API integration for real-time case status and order updates
My Role: Frontend & Voice Engineer — built the complete Android application, CameraX integration, multilingual voice I/O, and readability enforcement pipeline.
🏅 IEEE YESIST12 2026 · 6th Place — India Preliminary Round
Tell your phone what to do. In plain English. The agent figures out the rest.
What I Built:
- 🔄 ReAct Agent Loop — Reasoning + Acting loop where the model observes the screen state, plans the next action, executes via ADB, and re-evaluates — until the task is done or a max step limit is reached
- 📸 Screenshot Intelligence — Gemini Vision interprets the current screen state in natural language, allowing the agent to understand any app's UI without prior knowledge
- 🌲 Accessibility Tree Parsing — ADB dumps the view hierarchy; agent selects the most precise interaction target (by resource ID or bounds) rather than pixel-clicking
- 🔁 API Key Rotation — multi-key round-robin pool with cooldown tracking ensures near-zero rate-limit failures for long-running sessions
- 📡 WebSocket Progress Streaming — every agent thought, action, and result streamed to client in real-time; frontend shows a live agent log
- ♿ Accessibility Support — TalkBack + TTS controls baked in from day one; agent can operate the phone for users who need assistive technology
My Role: Sole developer — agent loop architecture, ADB bridge, screenshot pipeline, session management, streaming infrastructure.
Live AI idea evaluation at scale — built to survive 90+ concurrent users without breaking a sweat.
What I Built:
- 🏋️ Load-Resilient Architecture — multi-key API pool with round-robin balancing; fallback scoring kicks in automatically when primary LLM is rate-limited; no user-visible downtime
- 📥 Priority Queue System — request queue with priority lanes ensures fairness under peak load; no submission gets lost; latency stays predictable
- ⚡ Live Evaluation Feed — WebSocket-driven interface shows real-time scoring, feedback, and leaderboard updates as submissions come in
- 📊 Analytics Dashboard — per-participant scoring history, submission timelines, and event-wide leaderboard for coordinators
- 🔒 Session Management — JWT-secured evaluation sessions with anti-duplicate-submission guards
My Role: Sole developer — designed and shipped the entire platform for RVCE's flagship AI/ML event.
Impact: Served 90+ concurrent users stably during AI Odyssey 2025.
Personalized AI tutoring — built and shipped in under 24 hours under competition conditions.
What I Built:
- 🧭 Adaptive Study Paths — Gemini decomposes any topic into an optimally sequenced learning tree based on the user's stated goal and prior knowledge level
- 📈 Progress Intelligence — tracks completion rates, weak areas, and time-per-concept; resurfaces topics the user is struggling with
- ⚡ Real-Time Analytics — Firebase Realtime DB powers live instructor dashboard showing where every student is in the curriculum
- 🗣️ Conversational Tutoring — embedded AI tutor that answers follow-up questions with context from the user's current learning path, not generic answers
- 🏆 Gamified Checkpoints — quizzes generated per-topic by Gemini; adaptive difficulty based on past performance
My Role: Team Lead & Full-Stack Developer — led the team to 1st place in Karnataka and 3rd nationally.
🥇 1st Place Karnataka — Rewind & Recode 2025, IIIT Bhubaneswar
AI agents that execute financial workflows — every decision logged on-chain before execution.
What I Built:
- 🧩 4-Agent Orchestration — Planner decomposes intent → Executor selects transactions → Evaluator validates safety → Communicator reports outcome; all coordinated via LangChain
- 📝 Decision-First Architecture — every agent decision is hashed, timestamped, and anchored to the blockchain before execution; auditability is a design principle, not a feature
- 🔐 ERC-4337 Smart Accounts — Safe SDK-backed programmable wallets with configurable spending limits, emergency pause, and multi-sig support
- 📊 Live Telemetry Dashboard — WebSocket-powered real-time agent state, transaction history, and on-chain proof viewer
- 🛡️ AES-256 Key Management — private keys encrypted at rest; no plaintext exposure in memory during agent execution
- 🌐 Multi-Network — deployed across Ethereum Sepolia, Polygon Amoy, and Base Goerli testnets
My Role: Full-Stack Developer — agent pipeline, smart contract integration, blockchain audit logging, live dashboard.
| Project | What It Does | Stack |
|---|---|---|
| 🛡️ AllerSafe | Allergen detection from food labels via hybrid KG + DistilBERT · 96.7% macro F1 · 4.7s camera-to-result · IEEE & Scopus Published | FastAPI Flutter spaCy DistilBERT Supabase |
| 📚 EduSynth | Full educational content pipeline: topic → animated video lecture + mind map + PDF + PPTX · 3rd National, Rewind & Recode | Next.js FastAPI Gemini ElevenLabs Manim |
| 🛡️ Sentinel Orchestrator | Decentralised AI-agent threat detection swarm for Cardano DeFi · Masumi + Hydra L2 + Midnight ZK | Haskell Python CrewAI Cardano IPFS |
| 🔗 NeuraMark | Blockchain content authentication with semantic fingerprinting · ERC-721 NFT provenance + ChromaDB similarity | Next.js Solidity Hardhat ChromaDB IPFS |
| ⚙️ NexaKernel | Bare-metal OS kernel in C + x86 ASM · custom bootloader, paging, round-robin scheduler, VGA driver | C x86 Assembly QEMU GCC |
| 👁️ Glimpse3D | 3D reconstruction from 2D images · SfM, monocular depth CNN, mesh generation, texture mapping | Python PyTorch OpenCV SfM |
| 🏆 Achievement | 📍 Venue | 📊 Impact |
|---|---|---|
| 🥈 2nd Place — VISISONICS AI'26 | MIT Bangalore × Hitachi India R&D | 20-hour build · 3-person team · evaluated by Hitachi engineers |
| 🏅 6th Place India — IEEE YESIST12 2026 | International IEEE Student Competition | India Preliminary Round · Frontend & Voice Engineer |
| 📄 Research Publication — NQComp-2026 | IEEE & Scopus Indexed · Manipal Bengaluru | AllerSafe · 96.7% macro F1 · 847-node knowledge graph |
| 🥇 1st Karnataka — Rewind & Recode 2025 | IIIT Bhubaneswar · D³ Tech Fest | LearnMate AI · Team Lead |
| 🥉 3rd National — Rewind & Recode 2025 | National Finals | EduSynth · Full-stack development |
| 🎤 Bootcamp Coordinator | Coding Club RVCE 2026 | 130+ first-year engineers · 9 sessions · MongoDB track lead |
| ⚡ AI Odyssey 2025 — Event Lead | RVCE AI/ML Flagship Event | MindForge powered 90+ concurrent users live |
|
🧠 Context-Aware AI Designing LLM pipelines that maintain coherent multi-turn state without hallucinating prior context. Applied in INDRA's narration engine and Android AI Agent's ReAct loop. |
🤝 Multi-Agent Coordination Task decomposition, inter-agent trust, and emergent behaviour in LangChain / CrewAI orchestration — INDRA (7-layer pipeline), WalletMind (4-agent system), Sentinel Orchestrator. |
📱 Edge-First Intelligence On-device inference for low-latency, privacy-preserving mobile apps using TFLite + ONNX. NyayaSetu's offline-first legal pipeline is the proving ground. |
Open to: AI/ML Research · Open-Source · Hackathons · Startups · Systems Programming · Anything ambitious
"Build systems that reason. Not tools that respond."
$ shutdown --graceful
· Saving session state... ✓
· Flushing memory buffers... ✓
· Terminating active processes... ✓
· Archiving telemetry logs... ✓
Connection terminated.
System entering standby. █