shaafsalman

Shaaf Salman

ML/AI engineer working on medical vision-language models — chest X-ray report generation, clinical reasoning fine-tunes, and LLM production observability tooling.

Affiliated with NeuroCare.AI and SaveLife.AI. Writing at medium.com/@ishaafsalman.

🩻 Chest X-Ray (CXR) Vision-Language Models

Fine-tunes of Qwen3.5-9B for chest radiograph report generation, plus supporting object-detection and segmentation baselines.

🧬 Medical Reasoning Models

📊 Chest X-Ray Datasets & Evaluations

🧪 Synthetic Data & Reasoning Traces

Several of these datasets were generated with Mimesis, a local vLLM-based synthetic training-data pipeline — see the writeup: Generating Synthetic Training Data at Scale with vLLM.

🔒 Internal / Company Data (private)

Production LLM observability datasets from internal services — audit logs, benchmark samples, and regression fixtures. Kept private due to internal service and clinical-adjacent content.

llm-audit-logs · llm-audit-logs-by-service · llm-bench-prompts · medcoding-prompts · medcoding-v3-test-cases · soap-reasoning-leaks

✍️ Writing

Selected posts from medium.com/@ishaafsalman:

• You Don't Always Need an LLM to Extract Entities
• You Might Not Need a Vector Database Again
• The Onion Stack: Debugging vLLM, FlashInfer JIT, TMA Descriptors, and a Broken HTTP Router
• Run Qwen3.5-9B with vLLM — Streaming, No Thinking Mode
• Quantizing Qwen3.5-9B to INT8: RTN vs AWQ on a Hybrid VLM Architecture
• What Actually Broke: A Qwen3.5 Medical Fine-Tuning Followup
• Running MedXpertQA on Custom Vision LLMs: A Practical Evaluation Guide
• Run Stanford MedHELM on Any vLLM Model with One Bash Script
• Generating Synthetic Training Data at Scale with vLLM: Building Mimesis
• Fine-Tuning Qwen3.5: MoE vs Dense

---

Profile last reorganized 2026-07-21. No Spaces or Collections currently exist on this account.