Software Developer specializing in mobile application engineering with Flutter & Dart, native Android (Kotlin), Python backends (FastAPI/Django), privacy-first On-Device AI, browser-based AR (MediaPipe + Three.js), and evaluation-driven RAG systems. Certified in App Development with Flutter & AI (Ostad) with demonstrated experience engineering memory-constrained apps with 0 MB background Flutter runtime overhead, low-latency on-device inference on 2GB–4GB Android hardware, hardware-backed encryption, real-time cloud architectures, and measurable LLM quality (faithfulness / hallucination evaluation with ablation studies).
- 📍 Location: Kaliyaganj, West Bengal, India
- 🎯 Open to: Entry-level Software Engineering Roles, Full-Stack / Mobile / AI Engineering & High-impact Collaborations.
- 📬 Get in touch: royaa7840@gmail.com • LinkedIn
JavaScript (ES Modules) • Three.js (WebGL) • MediaPipe PoseLandmarker • WebAudio API • Zero Build Step
- Your Body Is the Controller: Real-time full-body pose estimation drives gameplay — step sideways = lane change, arms up = jump, squat = slide, T-pose = hoverboard. GPU delegate with automatic CPU fallback, 100% client-side, no install.
- Robust Gesture Pipeline: 2-second height/distance calibration, exponential smoothing, hysteresis lane bands, and per-gesture cooldowns turn noisy landmark streams into reliable real-time game events — with a live gesture coach and graceful touch/keyboard fallback.
- Complete Game Product: 4 unlockable runners with gameplay perks, rotating missions, coin-bank progression, 3 difficulty tiers, a fully procedural WebAudio soundtrack (zero audio assets), and a responsive premium dashboard from phone to desktop.
Python • Hybrid BM25 + Dense (RRF) • LLM Reranking • Gemini API • LLM-as-Judge
- Self-Evaluating RAG: Citation-grounded generation with an automated evaluation pipeline measuring Faithfulness, Hallucination rate, Answer Relevancy, Context Precision/Recall, and Refusal accuracy — every claim audited against its source passage.
- Measured Results (12-question golden set): Faithfulness 1.000 and hallucination 0.000 across all retrieval variants; the ablation study exposes dense-only precision dropping to 0.908 on keyword-heavy queries and RRF fusion recovering it to 1.0; both unanswerable questions correctly refused — zero fabricated answers.
- Pure-Python Retrieval Core: Okapi BM25 + cosine dense search fused with Reciprocal Rank Fusion, listwise LLM re-ranking (relevance grades 0–3), disk-cached embeddings, and one-command 4-variant ablation reports with a per-claim audit trail.
Flutter (Dart 3) • Kotlin • Riverpod • SQLite (WAL) • AES-256-GCM • llama.cpp
- Zero-Overhead Architecture: Engineered an offline Android security app for 2GB–4GB RAM devices with 0 MB background Flutter engine overhead via native Kotlin background interceptor.
- Real-Time Threat Classifier: Built a high-speed Tier-1 security engine (~8 MB RAM, 8–14 ms latency) detecting phishing, OTP harvesting traps, and homograph character spoofing.
- Privacy & Security: Implemented hardware-backed AndroidKeyStore AES-256-GCM envelope encryption,
Isolate.run()batch decryption, pure Dart TextRank +llama.cppFFI summarization, and a 60 FPS GPU-optimized UI.
Flutter • Kotlin • Accessibility Service • Multi-Process :agent • llama.cpp
- Low-Resource Overlay HUD: Architected a floating copilot operating smoothly on 2GB Android Go hardware (<= 8 MB incremental PSS, <= 50 ms p95 screen text extraction).
- Play Store Compliant: Designed strict suggest-only execution (SafetyGate confirmation flow) and automated local PII redaction (cards, OTPs, IBANs).
- Multi-Mode AI Pipeline: Offline deterministic heuristics, isolated
:agentnativellama.cppinference, and optional BYOK cloud LLM integration.
Flutter • Dart • Firebase Auth • Cloud Firestore
- Zero-Cost Scalable Architecture: Private real-time messenger built to operate seamlessly on Firebase with Google Sign-In and unique
@usernameclaims. - Optimistic Real-Time Engine: Implemented optimistic sending (
hasPendingWrites), message editing, soft deletion, typing indicators, live presence, and lightweight split-ABI APKs (22.9 MB arm64-v8a).
Kotlin • Jetpack Compose • Material 3 • Room • SpeechRecognizer • Gemini API
- Diagnostic Agent: Engineered a bounded troubleshooting workflow (observe → diagnose → consent → verify) running live DNS/HTTPS probes with strict tool allowlists.
- Multilingual Voice Engine: Built multilingual speech dictation supporting Bengali, Hindi, English, and Banglish paired with a Gemini structured-output proxy.
Flutter • Go (Chi) • PostgreSQL (pgvector) • FastAPI • NATS JetStream • MinIO
- Modular Monolith & Microservices: Material 3 Flutter client integrated with high-performance Go backend and Python background workers for isolated document ingestion and vector semantic search (
pgvector).
- 🏆 App Development with Flutter & AI — Ostad (Intensive hands-on training covering Flutter UI architecture, reactive state management [Riverpod, Provider], Firebase, REST APIs, and On-Device/Cloud AI workflows).
- 🏫 Secondary Education (10th / Madhyamik) — Kaliyaganj Parbati Sundari High School (WBBSE)