// paper #01
ReAct: Synergizing Reasoning and Acting in Language Models
ICLR 2023
Interleaving reasoning traces with actions beats either alone on knowledge and decision tasks.
Read breakdown
Khasim Bin Saleh
AI/ML Engineer, Berlin, Germany
I build LLM and ML systems end to end, and I measure them before I trust them.
// focus
- Deep Learning
- LLMs and RAG
- Agent Reliability
- Real-Time Voice AI
- MLOps
- Robotics
// about
About
I have spent two years building ML and LLM systems end to end. The work runs from data pipelines and feature engineering to evaluation and Dockerized deployment with tests.
At Coschool I shipped an AI tutor to more than 1,000 beta users. Now I am building Klindok, offline GDPR-compliant clinical documentation software for the DACH market. It runs on my own small language model.
My MSc thesis builds an LLM agent that checks whether synthetic health data is safe to share. I look at usefulness and privacy under GDPR.
// current work
Current work
MSc thesis
In progressAn LLM agent for assessing the utility and privacy of synthetic health data under GDPR.
Klindok
BuildingOffline, GDPR-compliant clinical documentation for private practices in the DACH market.
SiteWeekly paper breakdowns
OngoingOne AI paper analysed every week.
German B1
In progressGerman, in progress toward B1.
// latest from the lab
Latest from the lab
// week 01
Comparing enterprise AI agent surveys: why 17% vs 65%
2026-09-28 to 2026-10-04
Compared four 2026 industry surveys on enterprise AI agent adoption, where reported uptake ran from 17% to 65%. The gap came from who each survey asked, how it defined an agent, and whether a vendor ran it.
Read the week// selected projects
Selected projects
Klindok
Building03/2026, Berlin
Offline, privacy-first AI medical documentation. GDPR-compliant Arztbriefe in under 30 seconds for DACH private practices.
- < 30s per Arztbrief
- 100% offline
- GDPR Art. 30 audit log
- SQLite
- Fernet
- bcrypt
- pytest
2023
Built and shipped a 0 to 1 AI tutor serving 1,000+ beta users.
- 1,000+ beta users
- 80%+ fewer off-topic answers
- +35% contextual accuracy
- -40% retrieval latency
- RAG
- Milvus
- GPT
07/2025
22 derived features from 8 sensor streams, 5 models compared on F1 under class imbalance, 14 pytest tests, and a one-command Docker setup.
- 22 features
- 5 models
- 14 pytest tests
- XGBoost
- Docker
- pytest
// skills
Skills
Programming
- Python
- SQL
- CUDA
- LaTeX
ML and Deep Learning
- PyTorch
- TensorFlow
- scikit-learn
- XGBoost
- Hugging Face Transformers
GenAI and LLM
- LangChain
- RAG
- OpenAI
- OpenRouter
- Gemini
- Azure Document Intelligence
- prompt engineering
- real-time speech (Deepgram STT, TTS)
Evaluation and Benchmarking
- evaluation harnesses
- rubric-based LLM scoring
- evidence-based model and prompt selection
- statistical testing (Diebold Mariano, Friedman)
- time-series forecasting
- ensemble methods
Computer Vision
- OpenCV
- YOLO
- ArUco marker tracking
- real-time object detection
Tools and Infrastructure
- Docker
- Git/GitHub
- Linux (Ubuntu)
- Postgres
- Prisma
- FastAPI
- Streamlit
- React
- Power BI
// experience
Experience
06/2026 to 10/2026 · Berlin
Research Assistant
SRH Berlin University of Applied Sciences
- Compared four 2026 industry surveys on enterprise AI agent adoption, where reported uptake ran from 17% to 65%, and traced the gap to who each survey asked, how it defined an agent, and whether a vendor ran it.
- Reviewed five agent design patterns (ReAct, Reflection, Planning, Tool use, Multi-agent), going back to the original ReAct and Reflexion papers, and matched them to live deployments in banking and healthcare, and to the MCP and A2A protocols.
03/2023 to 08/2023 · Hyderabad
Machine Learning Intern
Coschool
- Built and shipped a 0 to 1 AI tutor serving 1,000+ beta users.
- Designed evaluation pipelines to catch unsafe or off-topic outputs, and fine-tuned GPT models with guardrails, cutting incorrect or off-topic answers by 80%+ across 500+ test cases.
- Built a RAG pipeline that improved contextual accuracy by 35%+.
- Deployed semantic chunking and Milvus ingestion for 10K+ documents, cutting retrieval latency by 40%.
// beyond code
Beyond code
- Patent
Automatic and Internet-Controlled Mobile Pet Food Dispenser System
Application No. 202311073080, filed October 2023, co-inventor.
- Student Council
- Student Council candidate, SRH (2026).
- Interests
- Business and research reading
- Travel
- Philosophy
- Healthcare AI
- Languages
- Hindi (native), English (C1), German (B1, in progress)