Work

Roost

Open-source Postgres-backed job queue for Python. No Redis, no RabbitMQ — just your existing Postgres. Transactional enqueue, LISTEN/NOTIFY, crash recovery, and a live dashboard. 15,289 jobs/sec on a single async connection.

Agent-UniRAG

Open-source trainable LLM agent framework for unified single-hop and multi-hop question answering. Five-node LangGraph state machine with ChromaDB, FastAPI streaming, and a Next.js UI. Based on arXiv research.

KERAG

Implementation of Knowledge-Enhanced RAG (arXiv:2509.04716): combines SPARQL-based knowledge graphs with LLM relevance scoring and Chain-of-Thought reasoning for complex multi-hop QA. LangChain, LangGraph, multi-provider LLM support.

GameMind Agent

Conversational AI platform for game developers using LangGraph with PostgreSQL vector embeddings. Boosts idea recall accuracy by 65% through semantic memory retrieval and automated visualization. Built at Schoene Neue Kinder GmbH (Munich, remote).

HANRAG

Implementation of the HANRAG framework (arXiv:2509.09713v1) for robust retrieval-augmented generation under noisy, multi-hop query conditions. Python, RAG, LangChain, LangGraph.

Multi-Agent Slack Bot

Multi-agent system automating offer generation, lead analysis, and follow-ups via LangGraph with Slack, Gmail, and Jira. Reduced manual effort by 65% and response time by 50%. Schoene Neue Kinder GmbH (Munich, remote).

LDFA Anonymization

Fine-tuned Meta's SAM for German license plate detection in self-driving applications. Improved accuracy from 75% to 95%, reduced false positives by 15%. W&B, Flux Inpainting. FZI Forschungszentrum Informatik (Karlsruhe, remote).

ChatTuesday

Multimodal RAG chat system integrating LLMs with text-to-image and text-to-video diffusion models via FastAPI. Reduced generation latency by 30%, improved cross-modal coherence by 25% using pgvector. GagaMuller (Dublin, Ireland — remote from Islamabad).

NoiseVu

Deployed Wave2Vec2 sound classification on Raspberry Pi for edge-based crime prevention. Detection accuracy increased from 70% to 92%, latency reduced by 35% via quantization. Data Pilot (Lahore, Pakistan).