I'm Ravani Roshan, an AI-focused builder working at the edge of autonomous agents, developer tooling, and systems that make large language models useful outside demo videos.
My current work is centered on agent reliability, sandboxed coding systems, and computer-use automation — the kind of infrastructure agents need when they stop being cute chat boxes and start touching real tools, real budgets, and real machines.
I like building products that feel sharp, fast, and a little dangerous in the best way: practical systems with enough personality to make the terminal smirk back.
Backstop is an in-process reliability layer for AI SDK usage: backpressure, budget enforcement, retries, circuit breaking, and metrics before requests leave the application.
It explores a simple but important question: what if multi-agent systems had brakes before they burned money, rate limits, or everyone's patience?
- Transport-layer control for AI SDK calls
- Budget isolation for multi-agent coding architectures
- Circuit breakers, priority admission, retry logic, and observability hooks
- Python-first implementation with OpenAI and Anthropic-oriented workflows
Nikki is a hermetic multi-agent coding system where isolated AI agents independently plan, code, test, and review inside Docker sandboxes, then hand back a reviewable Git branch.
Think: fire-and-forget coding agents, but with containment. Because giving five agents shell access without guardrails is how you accidentally summon DevOps Voldemort.
- Rust-based CLI and orchestration
- Docker-backed isolation
- Bring-your-own-key agent workflows
- Multi-agent planning, implementation, testing, and review
| Project | What it is | Stack / Focus |
|---|---|---|
backstop |
AI SDK backpressure, budgets, retries, circuit breaking, and metrics | Python, AI infra, agent reliability |
niki |
Hermetic multi-agent coding system with Docker sandboxes | Rust, Docker, CLI, AI agents |
phantom |
Invisible background-mode computer-use agent for Windows | Rust, Windows automation |
RavaniRoshanSingh |
Portfolio site and personal launchpad | TypeScript, product design |
roshans-portfolio-site |
Mac-desktop-inspired personal portfolio | React, TypeScript, Vite, Tailwind |
REI |
Repository Evolution Intelligence for predicting downstream file impact | Python, static analysis |
llm-finetuning-toolkit |
Toolkit for synthesizing data, aligning, and deploying custom AI agents | Python, SLMs, DPO, agent tooling |
mnemo-cognitive-telemetry |
Cognitive observability for memory-augmented LLM agents | Python, agent memory, observability |
- 💼 LinkedIn: Ravani Roshan
- 🌐 Portfolio: roshans-portfolio-site.vercel.app
- 🧪 GitHub: RavaniRoshan
- 🐦 X: @RoshanAIs
- 📧 Email:
ravaniroshansingh@gmail.com
agent status: awake
coffee level: suspiciously high
build target: useful systems, not shiny demos
I build agents, sandboxes, reliability layers, and tools that make computers do the boring parts.
Occasionally, the computers complain. That means it's working.