$ ls ~/projects

// 5 builds — systems, backend, GPU compute, and AI.

›Production food-delivery backend: co-built 340+ REST routes across 36 domains, scaling 300 → 1,500+ users.
›Razorpay payments with signature verification, idempotency & reconciliation — zero failures across 2,500+ orders.
›Cut cold catalogue load 4.5s → 560ms (8x) via a versioned Redis snapshot API + gzip + MMKV; cloud cost down 66% (managed RDS/ElastiCache → self-hosted PostgreSQL/Redis on EC2).

›End-to-end GraphRAG pipeline - crawling, knowledge-graph construction, BFS traversal, MiniLM embeddings, HDF5 vectors, Gemini generation.
›Graph-based context expansion resolves multi-hop queries across disconnected pages that flat RAG misses.
›Served through a Node.js / Next.js interface over the Python retrieval + generation backend.

›Haptic navigation headband for blind / low-vision users - Orbbec depth camera feeds a 27-cell grid driving 27 vibration motors in real time on-device, no phone for the reflex layer.
›On-device Qwen3-VL-4B + YOLOv11 on the Snapdragon Hexagon NPU for object/scene description at true depth-measured distance, spoken via Sarvam ASR/TTS in 11 Indian languages.
›Embedded firmware: STM32U585 → PCA9685 PWM over I2C for 27 motors, coordinated with an Arduino UNO Q (QRB2210 arm64 Linux) host over MsgPack-RPC.

›Cloud-native expense tracker: upload a receipt photo and financial data is auto-extracted via AWS Textract OCR.
›Auth with email verification and JWT via AWS Cognito, with per-user data isolation.
›Dashboard with pie charts, trend analysis, and expense tables; monthly email summaries over SNS.

›On-device image-processing app: 35 classic filters written from scratch in TypeScript across 7 categories (intensity, histogram, smoothing, sharpening, edges, segmentation, noise).
›Custom pipeline decodes JPEG to an RGBA byte array, runs pure-TS filters, and re-encodes on-device.
›Parameter sliders, before/after comparison, filter stacking, and per-filter 'how it works' explainers.