I build agents
and the harnesses
that hold them

Manthan manthan.sirune.tech ↗

AI-powered spaced repetition, built end to end

Upload a document; Manthan turns it into concept cards, tags them, works out which ideas depend on which, and shows you the whole structure as a 3D graph you can fly through. I own it — architecture, roadmap, billing, all of it.

knowledge mode

A roadmap view on a Sugiyama-style layered layout — topological leveling, barycenter crossing-reduction, coordinate straightening — ranked by a custom leverage metric: downstream value unlocked per unit of reading cost.

engine

A stateless Python/FastAPI microservice using DSPy for card structuring and tagging, behind a Kotlin/Spring Boot API. SM-2 scheduling, real-time card creation over chat, and a direct JSON/Markdown import path.

unglamorous half

Keycloak SSO across multiple realms, and subscription billing written from scratch — Razorpay and Dodo Payments, webhook-verified activation, prorated plan switching.

agent-native

An MCP server at api.sirune.tech/mcp — decks and cards can be built straight from Claude Code, Claude Desktop, or any MCP client.

DSPy · FastAPI · Kotlin/Spring Boot · Three.js · PostgreSQL · Keycloak · Cloud Run · MCP

VeriSolve

Rust-native agentic hardware-verification harness · SiRune

A 26-crate Rust workspace — agent loop, permission and sandbox runtime, tool execution. I contributed the Python side.

dspy bridge

A bounded Python stdio JSON-RPC worker exposing five tools to the Rust agent core, including UVM log analysis. Every dangerous capability — shell, file writes, EDA submission — stays in Rust.

into the harness

Wired worker launch into the TUI's turn-executor, added /agent <name> for mid-session switching, fixed transcript auto-follow during streaming, hardened the log analyzer's regex fallbacks.

Rust · Python · DSPy · JSON-RPC · TUI

Gridweave read the writeup →

An agent that writes React components and shows them to you instantly

A DSPy ReAct agent — Gemini, esbuild, SSE streaming — that generates, compiles, and renders components live. The interesting part is the self-optimizing prompt-evolution loop (GEPA), which improves output quality over iterations against an evaluation signal instead of leaning on one hand-tuned prompt. Local and trusted-network use only — it executes model-generated code.

DSPy · Gemini · esbuild · SSE · GEPA

memqi github ↗

Spaced repetition that lives inside your Obsidian vault

A Tauri/Rust desktop shell over a real SM-2 scheduling engine, parsing three flashcard syntaxes directly out of existing vault markdown — no separate database, no re-entry. Includes an LLM "note to cards" workflow with an edit/accept step before anything is written back.

Tauri · Rust · JavaScript · SM-2

rune-tui read the writeup →

A terminal client for RTL and testbench generation

About a thousand lines of Rust — ratatui, crossterm, tokio — with the render loop and the network loop fully decoupled over channels, so tokens stream in while the interface stays responsive. Generated code opens in your real $EDITOR before it goes anywhere.

Rust · ratatui · tokio · reqwest

concept-cards

Where Manthan started

A Rust TUI reader paired with a Python/DSPy document-to-flashcard converter and a self-optimizing generation prompt. Terminal-native, and the direct ancestor of Manthan's architecture.

Rust · Python · DSPy

EvalAssist github ↗

Grading handwritten answer sheets for government-school teachers

A collaborative project led by others. My contribution was rebuilding the OCR and grading backend — off a local model stack (TrOCR/PaddleOCR/EasyOCR + torch) onto an OpenRouter Gemini Vision pipeline — plus production deploy fixes across auth, routing, MongoDB TLS, and stale service-worker caching.

Python · Gemini Vision · OpenRouter · MongoDB