safzan pirani

agent engineer · india · remote

i build efficient tools for agents to help them do long-running tasks with minimal human intervention. i built fleet so agents can do tasks on remote machines without running on them directly, and ruddr for steerable live subagents that an orchestrator agent can manage and watch live. i currently do agentic ai and llm work at altuni labs, building hiring solutions for insideiim.

pi, langchain / langgraph, typescript, python, go, next.js, fastapi, postgres / supabase, cloudflare workers + tunnels, vllm / sglang / llama.cpp / whisper.cpp, comfyui, telegram bots, tailwind / vanilla css

software development engineer @ altuni labs @ insideiim ↗

sep 2025 – present

remote

  • —build agentic b2b hiring features: model selection, tool layer, infrastructure, and the operator ui
  • —ship llm workflows with streaming responses and tool use to production users
  • —fix agent reliability problems as they show up under real traffic

ai engineer intern @ radicalx

dec 2023 – feb 2024

remote

  • —built career-coaching tools with openai and vertex ai
  • —built conversational tools for resume analysis and interview preparation
  • —explored fine-tuned stable diffusion and llama models

fleet

jul 2026 – present

fleet lets agents manage windows, linux, and mac machines over ssh through a cli or mcp server. it runs commands across hosts, manages background jobs, transfers and edits files, captures screenshots, and reports host and gpu status. built with bun and typescript.

openports

feb 2026 – present

openports finds publicly exposed comfyui and ollama servers, checks them on a schedule, and sends change alerts through telegram. the dashboard includes saved searches, watch lists, model listings, and change history. built with fastapi, sqlite, and react.

ruddr

jul 2026 – present

ruddr steers a running codex, claude code, opencode, or pi turn mid-flight. it owns the provider connection, writes state and events to plain files, and exposes a local socket so an orchestrator agent or a person can redirect the turn instead of killing it. an opentui dashboard shows every session live. built with bun and typescript.

see

aug 2026 – present

see converts image layout and text to ascii for text-only models. it can use a vision model when the ascii output needs more detail. built with bun and typescript, with cli and mcp interfaces.

voicewin

dec 2025 – present

voicewin records speech with a global hotkey, transcribes it through whisper or deepgram, and pastes the text. built for windows with c# and .net.

a comfyui node compares a flux kontext edit with the original image and merges only the changed regions. it preserves the original pixels elsewhere.

cliproxyapi-cursor (fork)

feb 2026 – may 2026

patched router-for-me/CLIProxyAPI to translate cursor's anthropic tool calls into gemini's functionCall format. added tool-name sanitization, image support, and unit tests.

local rag system

nov 2023 – mar 2024

an offline question-answering system uses ollama and llamaindex to search multiple data sources and stream responses.

license plate recognition

aug – oct 2023

detects indian license plates in video with yolov8 and reads the plate numbers with easyocr.