Engineering Case Studies
Deep-dives into the architecture, engineering trade-offs, and production infrastructure of the AI systems we design and build.
Sharathi: 24/7 AI Agent That Auto Replies to Facebook Messenger
How we built Sharathi: an open-source AI agent that answers Facebook Messenger the moment a message lands, streams every conversation to a live dashboard, and lets a human tune the agent's tone and knowledge without touching code.
Conversation Platform: Live Chat and an AI Bot in One Widget
How we built a two-product SaaS — human livechat and a RAG AI chatbot — on one schema, one dashboard, and one 6.7 kB embeddable widget, in 12 weeks with 4 contributors.
PhotoFox AI: One Snapshot to a Full Studio Shoot
How we built PhotoFox AI: a Nuxt 3 studio and a Node/Express API that orchestrate six generative models behind one contract, train private per-subject LoRAs, run GPU jobs on async queues with automatic fallback, and meter every credit in deterministic code.
HeyLuna: A Personal AI Assistant That Takes Action
How we built HeyLuna before AI agents existed: real-time voice, 14+ tool functions across Gmail, Calendar, and Slack, vector memory, and OAuth integrations — at the start of the AI era.
Vocalo: Ultra-Low Latency Voice AI
How we built Vocalo's real-time voice agent: migrating from a high-latency, multi-hop WebSocket pipeline (Azure & ElevenLabs TTS) to a unified WebRTC streaming pipeline using Pipecat and Daily.co, bringing response latency under 1.5s.
InvoiceAgent: AI Document Intelligence Engine
How we built InvoiceAgent's backend: a redundant dual-extraction OCR pipeline, automated LLM schema enforcement, and a natural-language search layer to achieve financial-grade data accuracy on autopilot.
ApexGolf: An AI Golf Coach Grounded in Reality
How we built ApexGolf's coaching engine: expert-authored content and a deterministic rules layer handle every safety-critical decision, while a constrained Claude selector picks the plan — it selects, it never invents.
Nongor: Local-First Knowledge Base
How we built Nongor: a high-accuracy RAG engine that runs retrieval, embedding, and reranking entirely on-device, fuses dense and sparse search, and validates every citation against its source — so answers are trustworthy and no document data leaves the machine.
Quizmaker: Any Document to a Graded Quiz
How we built Quizmaker: a full-stack platform that turns typed text, PDFs, and images into schema-validated quizzes with GPT-4o-mini, then shares, auto-scores, and exports every attempt through deterministic code the model never touches.
SketchToImage: Turn Any Sketch Into Art
How we turned a weekend scribble-to-image experiment into a production creative studio: a real drawing canvas, scribble-conditioned generation across 20 styles, an async GPU job pipeline, plus upscale, image-to-video, credits and payments around it.
Offline AI Chat: Fully On-Device, Privacy-First AI
How we built a React Native chat app where chat inference, voice input, and voice output all run as local model inference on the phone — no server involved at any point after the first model download, verified in airplane mode.