Notes from shipping production AI systems, LLM applications, and Django backends. The architecture decisions, the trade-offs, and what actually happened in production.
How Nepal's #1 Bookstore Syncs Inventory in Real Time Tracking stock across our branches used to mean manual entry, errors, and hours of order processing. We built a hub-and-spoke system around Repo-v1 that syncs inventory in real time and routes orders to whichever branch has stock.
Building Robots That Remember Instead of Getting Stuck Reactive robots drive into the same corner forever. We built four progressively smarter AI agents, and adding memory plus a state machine produced an adaptive system that navigates complex mazes 4.5x better than the reactive one.
Killing Manual Data Entry with Vision and Language Models Typing in thousands of book records by hand was soul-crushing, so we automated it with Gemini Vision and LangGraph: a pipeline that pulls structured data off book covers and validates it against our genre taxonomy.
When Book Search Became a Conversation Standard search choked on queries like 'books about overcoming grief'. We built a semantic search engine with LLMs and WebSockets that understands what readers actually mean and streams back answers in real time.
Evolving the Best Schedule for Mobile Hardware Scheduling neural networks on heterogeneous mobile chips is NP-hard, so we didn't try to solve it exactly. Based on the paper by DUSEOK KANG, JINWOO OH, JONGWOO CHOI, YOUNGMIN YI And SOONHOI HA, we built a genetic algorithm framework that evolves schedules, trading battery life against execution speed.
Reading the Unreadable with Microscope Slide OCR Microscope slide labels are tiny, faded, and rotated, and standard OCR can't read them. We paired OpenCV preprocessing with PaddleOCR and hit 95% accuracy on labels Tesseract gave up on.
How ERP v2 Became the Backbone of the Business Our monolithic ERP broke something every time we touched it. We rebuilt it as a modular, event-driven system on Django signals and SDKs, so features scale independently and third-party chaos stays out of the core.
Opening 65k Books to Every Developer with BM Open API Most book data sits behind paywalls and API keys. We built a free public REST API instead: 65,000+ books, sub-second search on a pre-computed SQLite index, no signup required.
How CoverCraft AI Writes 10k Descriptions Without Getting Banned 10,000 books, zero descriptions, and an API that bans you past 40 requests per minute. We built an async pipeline with a sliding-window rate limiter and finished the whole catalog in about 6 hours.
An AI Legal Assistant That Goes From Chat to Action Most chatbots are glorified FAQ pages. We built an agentic AI on FastAPI and WhatsApp that actually books appointments and hands the lawyer a qualified, scheduled lead instead of a raw inquiry.
Automating Blog Publishing with Django Signals We kept forgetting to trigger the n8n workflow after publishing a post, so we wired Django signals to fire it automatically. One status change in the admin and the whole distribution pipeline runs itself.
The Backend Behind Freedom Adventures Adventure tourism runs on permits, guides, vehicles, and health forms, and spreadsheets can't hold it together. We built a booking engine with structured itinerary modeling and real-time resource management to carry the load.
Running an Ultra-Marathon's Logistics on One System Race registration in the Himalayas used to mean Excel merges and duplicate bib numbers. We built one event platform with an on-site POS and live leaderboards, so every runner on the start line actually exists in the database.
Building a Multi-Tenant ERP from Scratch for Many Businesses Multi-tenancy is hard to get right. We built a schema-based architecture that serves hundreds of businesses from a single database, with hard data isolation on shared infrastructure.
Solving Trust Issues in DegenHouse with Smart Contracts Blockchain betting is usually slow, expensive, or both. We built a hybrid platform: Web2 speed for gameplay, Web3 settlement on Arbitrum, and Chainlink VRF for provable fairness.
One Loyalty Program Across Every Store Flat cashback rewards low-margin sales you lose money on. We built a centralized loyalty microservice that calculates rewards from actual profit and unifies customers across three different ERPs.
Automating Tax Compliance for an Entire Industry Nepal's tax API goes down a lot, and every invoice legally has to sync with it. We built a store-and-forward relay with exponential backoff and idempotency keys so sales never stop and nothing files twice.
How GweiStation Watches Gas Prices So You Don't Have To Gas fees swing from $2 to $50 in minutes, and mistiming a transaction costs real money. We built a real-time gas monitor on Redis TimeSeries and Celery that alerts users when it's cheap to transact.
One Source of Truth for a Multi-Vendor Marketplace Book data was scattered across vendors that couldn't agree on formats, prices, or even what a book is called. We built a central repository engine with Elasticsearch and priority-based conflict resolution to merge millions of records into one source of truth.
How Bizmandala Turned ERP into a Subscription Business Onboarding a tenant used to mean a developer running SQL scripts by hand. Bizmandala is the internal OS we built to automate provisioning, billing, and feature gating, turning our ERP into a real SaaS business.
Cleaning Up My Crypto Wallet with Dust Sweeper Crypto dust is money you can't see: tiny token balances your wallet UI hides. We built a multi-chain scanner that finds them and checks real-time gas costs to tell you which ones are worth moving.
How Reddit Shelf Signal Turns Noise into Leads Buyers on Reddit ask 'where can I buy this?' and move on within hours. We built a Python bot with PRAW that watches for those phrases in context and pings Discord the moment one shows up.
Teaching a Machine to Speak Nepali Existing Nepali TTS sounded robotic and choked on Devanagari. We trained a Tacotron model (building on existing implementations) with a Griffin-Lim vocoder that learned the script's quirks, schwa deletion included, straight from the data.
Crushing 15GB of JPEGs with a Custom Image Compressor 65,000 high-res images were killing our page loads. A two-thread pipeline built on Guetzli and Pngquant shrank the whole catalog by 60% overnight, with no visible quality loss.