Rohit Patil

The English tutor you can't catch teaching.

You add the words and phrases you want to use fluently, give the AI a persona, and just chat — by text or by voice. Underneath, a hidden engine weaves your words in and scores how well you use them, and it keeps a living memory of your life, growing more familiar the more you talk.

• Invisible pedagogy — its hidden mission is your vocabulary, but it never reveals the machinery. It just feels like a friend who picks good topics.
• A judge that never interrupts — a second model silently scores every message on how well you used the word, and picks what you practice next by spaced repetition.
• Memory it curates itself — living files on who you are, your life events, and your history with each persona, edited by the agent mid-conversation through tool calls.
• Talk to it for real — duplex voice calls over the Gemini Live API, scored live as you speak.
• A voice that runs the app — a second assistant reads your live app state and acts on it, creating personas, adding words, switching models, hands-free.
• Real product plumbing — Google sign-in, per-user isolation, and each user's own Gemini key encrypted at rest.
• Web and Android, one codebase — the same React app shipped native through Capacitor, with native sign-in and a shared session.

AI agents that think privately and speak socially.

A multi-agent framework built to make a discussion feel like a real room of people — not a tidy sequence of function calls.

• A private inner life — each agent has hidden thoughts, memories, perceptions of the others, and its own motivations.
• The floor is earned — before speaking, an agent decides whether it even wants to, then competes by urgency and social balance. No fixed turn-taking.
• Public ≠ private — what an agent says can deliberately differ from what it's really thinking.
• Mix and match — OpenAI, Anthropic and Google models, so different agents behave like different people.

A second brain that remembers who you are — not what you saved.

A voice-first AI agent inspired by Andrej Karpathy's "LLM Wiki" pattern. It quietly compiles a living snapshot of your life, relationships and focus through ordinary conversation.

• An identity.md working memory split into four layers — Core Identity, Life Phase, Events, Active Focus.
• Time-based decay — entries are timestamped, and unreinforced memories sink deeper, the way we actually forget.
• A synaptic knowledge graph — people and projects as markdown files with permanent 14-digit IDs, linked like Obsidian notes.
• Talk or type — browser voice mode via Gemini Live, or local tools through MCP.

A council of AI experts, not a single chatbot.

For the hardest decisions, one viewpoint isn't enough. Manthan convenes a panel of expert personas, has each reason through your problem independently, then churns their answers into one stronger verdict.

• Each expert, its best model — an engineer on Gemini, an ethicist on Claude, an economist on GPT. One conversation, many minds, each playing to its strength.
• Blind first, then a real debate — every expert answers without seeing the others (no groupthink), then an optional round two where they read each other and revise or defend.
• Synthesis that decides — a guide agent reads the whole panel and writes one decisive answer for you, surfacing disagreement only where it changes what you should do.
• Local-first, your own keys — no accounts, no cloud; everything runs on your machine with every token and dollar tracked.

An MCP server that drops AI straight into your codebase.

A "vibe coding" server with 15+ tools for file operations, code analysis, shell commands and project exploration — packaged as a one-click desktop extension that colleagues and senior devs actually adopted.

• 20–50% less development time through AI-driven generation, debugging and documentation.
• Syntax-aware editing and smart file operations, not blind text replacement.
• Sandboxed execution with comprehensive error handling for secure AI workflows.
• Shipped as a .dxt desktop extension — install in one click, no setup.

A structured, human-like memory that any agent can plug into.

An MCP server that swaps unstructured text dumps for a rigid four-level memory tree — so an agent files what it learns the way an engineer would.

• Four levels — Core Identity, Projects, Tasks, and granular Micro-logs, all backed by SQLite.
• Hybrid search — 60% semantic meaning blended with 40% keyword ranking, so it finds both exact facts and broad ideas.
• Two-step retrieval — skim the summaries first, pull deep logs only when needed, and the context window never bloats.
• Nothing forgotten by accident — full history is preserved.