Every project, graded honestly.
What is in daily use, what we have measured, what works but is not benchmarked yet, and what is still an idea. We publish a number only when we can reproduce it from a log.
ALL PROJECTS · 9
IN USERK OpsA portal and WhatsApp agent that run an institute's front office: absence alerts, fee reminders, receipts and replies in everyday Hinglish.OPERATIONS AUTOMATION73/73 checks · 4 AI fallbacks · in use since Oct 2026MEASUREDKeepsoundA voice memory that never leaves the phone. Ask a hazy question; hear the answer in your own recorded voice.PRIVATE AI SEARCH · ON-DEVICE10/10 in top 3 · 5–13 ms per search · no networkIN USENexus RAGLong-term memory for AI coding agents, served locally over MCP by small specialist models instead of one big one.AGENT MEMORY · MCP505 starts · 111 queries · 82 ingests (Mar–May 2026) · latency re-run pendingMEASUREDReal-time ML pipelineA Python feature builder and a C++ model runner sharing memory at 100 Hz. The pipeline held; the first models did not, and we wrote up why.STREAMING DATA · PYTHON + C++24 h at 100 Hz, zero crashes · spikes cut 2,075 → 111 msPROTOTYPENeoGravityOne command that gives an AI coding assistant a persistent workspace: workflows, skills, tasks and research notes it can read every session.DEVELOPER TOOLING · NPMPublished on npm · @vibcoder/neogravity 4.0.0PROTOTYPENeoIndex AEOA store app that turns a product catalogue into a dense, factual document AI answer engines can read and cite.ANSWER-ENGINE OPTIMISATIONWorking prototype · not deployedPROTOTYPEHinglish translation pipelineAn LLM pipeline that produced a faithful bilingual English + Hinglish edition of a 136-page text, sentence by sentence.HINDI & HINGLISH AI3 versions built · style rules that stopped the driftEXPLORINGRAMUAn experiment in a long-running autonomous agent. Its biggest output was a set of guardrails, written after it broke itself.AUTONOMOUS AGENTS45-hour outage → hard rules · our own review: "ordinary RAG"EXPLORINGState space modelsTesting Mamba-style models on long data streams. The first GPU run diverged on un-normalised inputs — a lesson now built into our checklist.RESEARCH · MAMBA / SSM1 GPU run · hidden state hit ±10¹⁵ · next design gated
RK Ops was built for RK Institute of Learning, which our founder also runs. Client names appear here only with written permission. Demos and hackathon builds are not counted.