AI Engineer for Medical Data Extraction

@sajjadtaghvaeifr

Hello, I trust you're doing well. I am well experienced in machine learning algorithms, with nearly a decade of hands-on practice. My expertise lies in developing various artificial intelligence algorithms, including the one you require, using Python, and similar tools. I have worked with pytorch, and tensorflow to develop DL models, .I hold a doctorate from Tohoku University and have a number of publications in the same subject. My portfolio, which showcases my past work, is available for your review. Your project piqued my interest, and I would be delighted to be part of it. Let's connect to discuss in detail. Warm regards. please check my portfolio link: https://www.freelancer.com/u/sajjadtaghvaeifr

@ShaikhAneesa

Hi there, I understand you need a research-grade clinical NLP pipeline that can run open-weight LLMs entirely locally, extract structured information from pediatric congenital heart surgery records, provide evidence spans for every extraction, abstain when information is not documented, and reliably assign RACHS-1 and STS-EACTS categories. I’m confident I can build this as a reproducible evaluation framework rather than a simple LLM extraction script. My approach is to first establish the clinical schema and reference-standard format, then build a local Hugging Face/PyTorch inference pipeline with vLLM and appropriate quantisation. I’ll compare zero-shot, few-shot, and retrieval-based few-shot prompting using constrained JSON outputs, evidence-span linking, and not_documented handling. A second-stage reasoning pipeline will apply the risk-classification rules separately from extraction, making errors traceable to either extraction or scoring. I’ll evaluate models using accuracy, precision, recall, F1, weighted Cohen’s kappa, hallucination/confident-error rates, and detailed error analysis. I’ll also benchmark GPU memory, inference speed, document processing time, and computational cost while versioning models, prompts, configurations, and results through Git for full reproducibility. Do you already have the clinician-annotated reference-standard dataset and target GPU environment? I’m ready to start immediately. Warm Regards, Aneesa.

@koolvishruth

Hi Sir, I am AI engineer with 7 years of experience.I can work on your project as mentioned.I would like to connect through chat to know more in detail on requirement

@corpmember29

Hi, I’m interested in supporting your clinical NLP/LLM research project. I have 11+ years of experience in Python, ML/AI, NLP, PyTorch, Hugging Face, and LLM-based solutions. I can build the complete local LLM pipeline including open-weight model deployment with vLLM, quantisation, structured JSON extraction, evidence-span validation, abstention/not-documented handling, RACHS-1 and STS-EACTS reasoning, and comprehensive model evaluation. I can also handle zero-shot/few-shot comparisons, error analysis, GPU/inference benchmarking, reproducible experiments, and methodology documentation for the research manuscript. I’m available to discuss the dataset, target models, and research methodology and can start immediately. Thanks Christina

@shadabkhan92

With my extensive experience in Artificial Intelligence and Machine Learning, I believe I am the ideal candidate to spearhead your Clinical NLP project. My proficiency not only covers Python, Natural Language Processing, and Hugging Face Transformers, but also includes strategic prompt engineering for structured clinical information extraction - an integral task for your project. I have successfully built end-to-end clinical NLP/LLM extraction pipelines using frameworks like PyTorch and vLLM; this skill is directly applicable to deploying and testing open-weight LLMs locally as you require, without any data being sent to external APIs. You'll also benefit from my commendable track record in model evaluation and performance measurement. Evaluating the accuracy, precision, recall, F1-score, weighted Cohen’s kappa, hallucination rate, confident error rate of our models was part and parcel of our previous projects. Moreover,I bring a unique blend of skills that include deep understanding of hardware - both on PCB design and firmware - as well as software proficiency in React, Flutter, Django and Node. This expertise is valuable in ensuring the deployment of AI systems precisely where they are needed.

@TeamGTW

Hi, I’ve carefully read your requirements and understand that the key challenge is accurate clinical information extraction from medical notes using locally deployed open-weight LLMs, without sending sensitive data to external APIs. I can build the pipeline using Python, Hugging Face, PyTorch and vLLM, with: 1) Structured JSON extraction with evidence spans 2) not_documented / abstention handling to reduce hallucinations 3) Zero-shot, few-shot and retrieval-based prompting 4) Local model and quantization benchmarking 5) Separate RACHS-1 / STS-EACTS reasoning stage 6) Precision, recall, F1, kappa and hallucination evaluation 7) Reproducible experiments, prompts and model configurations My approach would be to first establish the extraction schema and evaluation set, then benchmark suitable open-weight models and optimize accuracy, reliability and inference performance. I understand this is clinical NLP/information extraction rather than a standard classification project, which is the most important part of your requirement. If you share the anonymized sample documents and reference-standard format, I can suggest the initial pipeline and model setup. Best Regards, Nikita Gupta

@Muhammadzeesha59

Hello Sir/MAM I am a Skilled Full Stack Developer. Having rich experience in Java , C++ , C , C# , Python , Eclipse , Sql , Mysql , .Net ,Oracle , Object Oriented Programming , Data Structure , Algorithms, Linux , Windows , Cloud , Azure , Ubuntu , OpenAI , Desktop Applications. Web Development I have a perfect grip on “Artificial Intelligence” “Automation” , and work in “Machine Learning” Deep Learning “Computer Vision ” Object Detection”. My track record as demonstrated in my 100% job completion and 5-star review rating showcases My ability to deliver exceptional results on time and with utmost quality I believe that my skill set makes me the ideal candidate for this project Please come on chat we will discuss more about this I will be waiting for your reply . Thanks and Best Regards

@ImranAli01

Hello, I got that you need a local open-weight LLM pipeline for clinical information extraction, evidence-span validation, abstention, and RACHS-1/STS-EACTS risk assignment without exposing clinical data to external APIs. This is what I can help you with, let's chat. My approach is to build the pipeline in Python using PyTorch, Hugging Face Transformers and vLLM, benchmarking quantized models with structured JSON schemas and constrained generation. I’ll compare zero-shot, few-shot and retrieval-based prompting, require evidence spans for every extracted value, and implement not_documented abstention to control hallucinations. Outputs will then pass through a reproducible second-stage risk-scoring pipeline and clinician-reference evaluation using F1, kappa, hallucination and confident-error rates, alongside GPU memory and inference benchmarks. As final deliverables you will receive the local NLP/LLM pipeline, model configurations, prompts, structured schemas, evidence validation, risk categorization, evaluation metrics, error analysis, computational benchmarks, reproducible Git-ready code and methodology support for the manuscript. One thing I'd like to confirm before we start: which GPU environment and reference-standard dataset are available? Let's discuss the setup and get started. Best Regards, Imran

@emilesd617

Hi, I am an AI/ML engineer with 8 years of experience in software development. I am familiar with Python, PyTorch, Hugging Face Transformers, vLLM, NLP, open-weight LLMs, prompt engineering, structured JSON generation, quantization, Pandas, NumPy, and model evaluation. I can build the local clinical extraction pipeline, compare zero-shot and few-shot strategies, link each extracted field to supporting evidence, add abstention logic to reduce hallucinations, and evaluate the results with accuracy, precision, recall, F1, kappa, and error analysis. I'm an individual freelancer and can work in any time zone you want. Please contact me with the best time for a quick chat. Looking forward to discussing more details. Thanks.

@Aaleen110

The challenge here is making locally deployed open-weight LLMs perform accurate clinical data abstraction on specialized medical records, and I'll tackle that by fine-tuning a model like Llama 2 or Mistral on a curated dataset of pediatric congenital heart surgery records, using Python and Pytorch for the development. I’ll use Prompt Engineering to design specific prompts that guide the model to extract diagnoses, procedures, and surgical risk categories from operative notes, discharge summaries, echocardiography reports, and cardiac catheterisation reports. I will not use cloud-based APIs for this project. Cloud APIs introduce latency and also introduce data privacy concerns which are not suitable for sensitive medical data. I'd like to know if there's a preferred output format for the extracted structured clinical information, such as JSON or a CSV. If you hand this over, I’ll spend the first day or two setting up the local LLM environment and preparing the initial data ingestion pipeline. 8 reviews on here, everything delivered on time and on the agreed price so far, plus Preferred Freelancer status.

@mayr81

Great fit for deep LLM-driven clinical NLP: building a fully local extraction pipeline that turns pediatric congenital heart surgery narratives into structured JSON (diagnoses, procedures, operative variables, outcomes, RACHS-1 / STS-EACTS) with evidence-span grounding. Will deploy and benchmark open-weight models locally across multiple sizes, using Hugging Face Transformers + vLLM, including quantized (AWQ/GGUF where suitable) inference for practical GPU memory/speed constraints. Implementation will cover zero/few-shot and retrieval-based prompting, structured/constrained generation for valid outputs, and an abstention mechanism to reduce hallucinations, every extracted field linked to supporting text. Then a second-stage reasoning module will map extracted variables to surgical risk categories, followed by rigorous evaluation vs clinician reference standards (accuracy, precision/recall/F1, Cohen’s kappa, hallucination/confident error rates) plus error analysis and computational performance reporting for a manuscript-ready methodology section.

@rosscarter1011

Hello ⭐️, Clinical abstraction for congenital heart surgery fails not from poor prompting but from missing ontological grounding—RACHS-1 and STS-EACTS aren’t just labels, they’re rule-based hierarchies that LLMs hallucinate when treated as free-text classification. The real bottleneck is encoding those rules into constrained generation so the model can only output valid categories backed by explicit text spans. I’d build a two-stage pipeline: first extract raw procedures/diagnoses with abstention, then apply deterministic mapping to risk scores using validated clinical logic, not LLM reasoning. This separates extraction error from scoring error during evaluation. Do you already have the reference-standard dataset with clinician-annotated RACHS-1/STS-EACTS labels, or will defining that ground truth be part of the initial phase?

@rezamahi

I’m an AI/ML engineer with hands-on experience in Python, NLP, LLM pipelines, Hugging Face Transformers, PyTorch, and structured data extraction. I understand this project requires much more than a conventional classification model—the key challenge is reliable clinical information extraction, evidence grounding, abstention, and reproducible evaluation. I can build the pipeline to run open-weight models locally using frameworks such as Transformers/vLLM, compare zero-shot, few-shot, and retrieval-based prompting, and enforce structured JSON outputs with validation. I can also implement evidence spans for every extracted variable and a not_documented mechanism to reduce hallucinations. For evaluation, I can build the complete benchmarking workflow covering precision, recall, F1, accuracy, weighted Cohen’s kappa, hallucination/confident-error rates, error analysis, and inference performance including GPU memory and processing time. The second-stage RACHS-1 and STS-EACTS assignment can be implemented as a separate, auditable reasoning/rule layer rather than mixing risk classification with extraction. I’m comfortable working with quantized open-weight models and maintaining reproducible model versions, prompts, configurations, experiments, and results in Git.

@vishal1145

Your LLM pipeline will fail HIPAA compliance if you're not implementing proper de-identification before local inference, and your extraction accuracy will drop below 70% without a multi-stage reasoning architecture that separates entity extraction from clinical logic application. Quick questions - are you planning to run inference on consumer GPUs or do you have access to A100/H100 clusters? And have you already annotated a gold-standard dataset or will this need to be built during the project? Here is the architectural approach: - PYTORCH + HUGGING FACE: Deploy quantized Qwen2.5 or Phi-3-medium models using vLLM with AWQ compression to fit 70B-class reasoning into 24GB VRAM while maintaining >85% extraction accuracy on structured medical entities. - PROMPT ENGINEERING: Build a two-stage pipeline where stage 1 extracts diagnoses/procedures with evidence spans using constrained JSON generation, then stage 2 applies RACHS-1 logic with chain-of-thought prompting to reduce hallucination rate below 5%. - NLP EVALUATION: Implement Cohen's kappa calculation across inter-annotator agreement, track confident error rates per entity type, and build an evidence validation system that flags extractions without supporting text spans for clinician review. I've built similar clinical NLP systems for 2 healthcare AI startups that processed 500K+ patient records under HIPAA compliance. Let's schedule a technical call to review your annotation strategy and GPU infrastructure before you commit to a model architecture.

@toriquldev123

Hello There! I'm Md Toriqul Islam, and I'm excited to partner with you & I can dive into your project immediately. I'm an AI/ML engineer experienced in building clinical NLP pipelines with locally deployed open-weight LLMs for structured data extraction. I understand you need an end-to-end extraction pipeline that reads pediatric congenital heart surgery documents locally using open-weight models, produces structured JSON output with evidence-span validation and an abstention mechanism, assigns RACHS-1 and STS-EACTS risk categories, and is evaluated against clinician reference data with accuracy, kappa, and hallucination rate metrics. I've handled similar clinical NLP and LLM extraction projects before. I am skilled in Python, Hugging Face Transformers, vLLM, model quantization with AWQ/GGUF, and structured prompt engineering for information extraction. I can walk you through my approach to zero-shot versus few-shot comparisons and reproducible experiment tracking once we connect. I'm ready to start right away and happy to discuss any details. Looking forward to hearing from you. Best regards, Md Toriqul Islam

@starttechno

Your core challenge isn't classification — it's getting an open-weight model to read messy operative notes and echo/cath reports, pull only what's actually written, cite the sentence it came from, and then reason RACHS-1 and STS-EACTS on top of that. That's how I'd build it. Everything stays local: vLLM serving Qwen and Gemma/MedGemma variants (plus a quantised AWQ/GGUF run for the smaller-GPU comparison), no external APIs touching PHI. Extraction runs against a strict JSON schema enforced with constrained decoding (outlines/xgrammar), where every field carries a value, an evidence_span, and a not_documented option so abstention is a first-class answer rather than a guess. Evidence spans get verified by exact match back to the source text, so anything unsupported is flagged automatically as hallucination. Risk categories run as a separate second-stage reasoning step over the extracted structure, which lets error analysis cleanly separate extraction failures from rule-application failures. Evaluation harness reports accuracy, precision, recall, F1, weighted Cohen's kappa, hallucination and confident-error rates against your clinician reference standard, plus GPU memory, tokens/sec and per-document latency per model. All prompts, model revisions and configs versioned in Git so results are reproducible for the manuscript, and I'll draft the methods text with you. Muhammad Saad

@LeadForce1

Hi, I'm Luigi from Toronto. "Local LLM Pipeline with vLLM + Structured Extraction" - you need accurate clinical data extraction from surgical notes using open-weight models. I'll build an end-to-end pipeline using Hugging Face Transformers and vLLM for local inference. I'll deploy and compare models like Qwen or Gemma with AWQ/GGUF quantization. For extraction, I'll design zero-shot, few-shot, and retrieval-based prompts with structured JSON schemas. I'll implement evidence-span validation (linking each extracted value to supporting text) and a not_documented/abstention mechanism to reduce hallucinations. For risk scoring, I'll build a second-stage reasoning pipeline for RACHS-1 and STS-EACTS categories. I'll measure performance (accuracy, precision, recall, F1, kappa, hallucination rate), compute GPU memory/speed, and provide reproducible code. Do you have a sample dataset and GPU environment ready? Looking forward to working with you. Luigi Tady

@HamparProd

Hi, This project is a strong match for my work with Python, local LLM inference, structured extraction and evaluation pipelines. I would build the system as a reproducible two-stage pipeline rather than asking one model to extract clinical facts and assign surgical risk in a single prompt. Stage 1 will process local clinical documents and produce schema-constrained JSON containing each requested variable, confidence/abstention state and supporting evidence span. Unsupported fields will explicitly return not_documented rather than encouraging model inference. Stage 2 will consume the validated extracted variables and independently assign RACHS-1 / STS-EACTS categories, allowing extraction errors and risk-rule errors to be evaluated separately. I can implement local inference with PyTorch, Hugging Face/vLLM and appropriate quantisation, then compare zero-shot, few-shot and retrieval-based approaches using accuracy, precision, recall, F1, weighted kappa, hallucination/confident-error rates and computational metrics. All clinical processing remains local, with versioned prompts, schemas, model configurations and experiment outputs for reproducibility.

@sufyanasmat38

I have experience with Python, LLMs, NLP, Hugging Face, PyTorch, and local model deployment, and I’m excited to contribute to this clinical NLP research project. What local GPU hardware is available for model inference and evaluation? I can build the end-to-end extraction pipeline with structured JSON, evidence spans, abstention handling, prompt evaluation, and model comparisons while keeping clinical data fully local. I can start immediately and deliver reproducible experiments, evaluation metrics, error analysis, and well-documented code suitable for the research workflow.