AI Systems & Backend Engineer · Applied LLM, Hybrid RAG, On-Device ML
CS Undergrad @ RV University (AI/ML) · Bangalore, India
🎓 RV University · AI/ML · CGPA 7.8 · Class of 2027 · 💼 2 industry internships shipped · 🧩 116 LeetCode (64 Med · 13 Hard) · 🏆 Best Project Award · IEEE CCEM 2024
I design and ship applied AI systems — local-first LLM pipelines, hybrid retrieval, MCP tool servers, and on-device ML on Apple Silicon. Production systems, not notebooks.
- Local-first LLM systems. Dual-LLM topologies (cloud + on-device fallback), schema-constrained generation, prompt-injection defense, deterministic guardrails for write operations.
- Hybrid retrieval. Knowledge graphs (NetworkX) fused with dense vector indexes (FAISS / pgvector) for grounded RAG.
- Model Context Protocol (MCP). Building MCP servers that expose tools (DB, travel, web) to LLM clients.
- On-device ML on Apple Metal. PyTorch + LightGBM training and inference targeting M-series GPUs via Metal Performance Shaders.
- Production backends. Spring Boot · Flask · FastAPI. RBAC, Fernet credential encryption, snapshot-based undo.
- DATABASE-MANAGER — natural-language to SQL across 8 engines with a dual-LLM safety pipeline.
- routecraft — multi-modal transit predictor for Bengaluru on Apple Metal.
- TRAVEL-PLANNER-MCP — MCP server exposing planning tools to LLM clients.
- graph-rag — hybrid Knowledge Graph + FAISS retrieval over user documents.
Previously: AI Intern @ Broadrange AI (Chicago) · Android Developer @ Techpuram (Madurai) Education: B.Tech (Hons) CSE @ RV University · AI/ML Major · FinTech Minor · Class of 2027
| Layer | Tools |
|---|---|
| Languages | Python Java C++ Kotlin C SQL Solidity JavaScript |
| ML / DL | PyTorch TensorFlow Keras scikit-learn LightGBM LangChain |
| Inference | Ollama Groq Hugging Face Transformers Mistral Llama BERT |
| Retrieval | FAISS pgvector NetworkX Chroma |
| Backend | Spring Boot Flask FastAPI REST Microservices |
| Mobile | Jetpack Compose CameraX Android SDK |
| Data | PostgreSQL MySQL SQLite MSSQL Oracle MongoDB Cassandra Redis |
| Infra | Docker Git Linux macOS / Apple Metal |
DATABASE-MANAGER — Natural Language to SQL across 8 engines
Local AI-powered database management with a dual-LLM topology: Groq-hosted Mixtral for high-fidelity generation, Ollama-hosted Mistral as the fully-offline fallback. Same NL prompt → executable query against any of 8 engines: SQLite, MySQL, PostgreSQL, MSSQL, Oracle, MongoDB, Cassandra, Redis.
Safety pipeline: schema-aware prompting, AST-level SQL validation, human-in-the-loop review on writes, snapshot + one-click undo, RBAC, Fernet-encrypted credential storage.
Python Groq Ollama Mistral Fernet RBAC 8 DB engines
routecraft — Multi-modal transit ML for Bengaluru
Predicts and routes across Walk · Auto · Cab · BMTC bus · Metro using ML-driven traffic priors. LightGBM gradient-boosted regressor for ETA, PyTorch sequence model for live re-estimation. Trained and served on Apple M-series GPUs via Metal Performance Shaders — no CUDA, no cloud GPU.
Python LightGBM PyTorch Apple Metal Geospatial
graph-rag — Hybrid Knowledge Graph + Vector RAG
NetworkX-backed knowledge graph fused with FAISS dense retrieval, synthesised by a Groq-hosted LLM. Document ingestion handles PDF / TXT with intelligent chunking. D3.js force-directed visualisation of the live graph. Flask backend, glassmorphism UI.
NetworkX FAISS Groq D3.js Flask
LLM-SERVICE-MCP — Autonomous agent on Spring Boot
Microservice-based AI agent. BERT intent classifier routes prompts in real-time to one of three downstream services: RAG, Web Search, or Chat. Multi-step reasoning over tool calls. Spring Boot orchestration in Java.
Spring Boot BERT RAG Microservices Intent Classification
LEGAL-AI-LLM — Built at Broadrange AI, Chicago
Legal-domain LLM grounded on curated case-law datasets to minimise hallucination. RAG architecture with document ingestion + chunking. PromptGuard V2 in front of the model for prompt-injection defense.
LLM RAG PromptGuard V2 Document Ingestion
Brain-Tumor-Segmentation — Medical imaging
U-Net encoder-decoder architecture for semantic segmentation of brain tumors from MRI scans. Built for a Kaggle competition.
U-Net TensorFlow Medical Imaging
Image-Resolution-Enhancer — Real-ESRGAN fine-tune
Transfer-learned Real-ESRGAN on paired HR/LR face datasets with custom loss balancing.
| PSNR | SSIM |
|---|---|
| 29.15 → 33.95 dB | 0.8985 |
Real-ESRGAN PyTorch Transfer Learning GANs
PERSONAL_ASSISTANT — Fully offline voice agent on macOS
Siri-style voice assistant for M-series Macs. Local speech recognition, local LLM inference, multi-turn dialogue with session memory, lightweight floating UI. Zero cloud calls.
Speech Recognition Local LLM macOS Python
GEO_LOCATION_SAVER_APP — Production Android, 100+ Play Store installs
Location-tagged image capture with server-side anti-spoof validation against fake GPS sources. Real-time address overlay via Google Maps API.
Kotlin Jetpack Compose CameraX Google Maps API
MORE PROJECTS
| Project | What It Does |
|---|---|
| TRAVEL-PLANNER-MCP | MCP server exposing trip-planning tools to LLM clients |
| INTELLI-CHAT-AI | Smart Document and Web Assistant — RAG + LLM |
| Blockchain-Voting-System | On-chain voting — Solidity smart contracts |
| Movie_Recommendation_System | Moodix — TF-IDF + Cosine Similarity recommender |
| Currency-Exchange | Live FX conversion application |
| Fake-News-Check | Fake news classifier — ML + NLP |
| Bank_Statement_Organizer | FinTech statement parsing and categorisation |
| WEATHER_SONG_RECOMMENDATION | Weather-conditioned song recommender |
| Print-Quality-Dataset-Generator | Synthetic dataset generation for print-quality CV |
| DELHI-WEATHER-ANALYSIS | Time-series analysis of Delhi climate data |
Artificial Intelligence Intern — Broadrange AI · Chicago, Illinois
Jun 2025 — Aug 2025
Built a legal-domain LLM with RAG retrieval over curated case law. Document ingestion + chunking pipeline. PromptGuard V2 for prompt-injection defense. Reduced hallucination through tight retrieval grounding.
Android Native Developer — Techpuram Technology · Madurai, Tamil Nadu
Feb 2025 — Aug 2025
Shipped Geo GPS Camera (100+ Play Store installs). Jetpack Compose + Kotlin + Google Maps API. Server-side validation against GPS spoofing.
B.Tech (Hons) Computer Science — RV University, Bangalore 2023 — 2027
Major: AI and Machine Learning · Minor: FinTech · CGPA 7.8 / 10
Achievements
- 🏆 Best Project Award — Structural Innovation, RV University
- 📄 IEEE CCEM 2024 — paper presenter (GlucoSense: non-invasive saliva glucose biosensor)
- 🧩 LeetCode — 116 solved (39 Easy · 64 Medium · 13 Hard)
Certifications — AWS (ML Terminology · CLI · DevOps Testing) · IBM SkillsBuild (Big Data, Hadoop, Spark, Kubernetes & OpenShift) · Google Generative AI · NPTEL Programming in Modern C++ · Udemy Spring Boot
Local-first. Hybrid retrieval. Production backends.
LLM systems · MCP servers · On-device ML · Spring Boot · Android
Open to AI/ML · Backend · Android internships — let's build something. 📫 ayonaryan5@gmail.com · LinkedIn