ICU teams generate massive volumes of time-series EHR data. Early identification of deteriorating patients is manually demanding and inconsistent. The system flags high-risk patients using historical ICU trajectories with an explainable probability score clinicians can act on at 3am.

Data governance: the repo uses a synthetic data generator replicating MIMIC schema distributions. Anyone can clone and run the full pipeline without PhysioNet credentials — a design choice I now cite in interviews as proof of data governance awareness.

• AUC-ROC 0.978 and AP 0.892 on held-out cohort with strict patient-level splits
• Temporal windowing using 6–12 hour bins; leakage controls enforced at data splitting level
• Model calibration verified across risk deciles — not just raw AUC
• SHAP explanations validated by ICU clinicians, reducing black-box concerns for high-risk flags
• Prediction drift detection and automated retraining triggers in production monitoring
• p95 inference latency under 800ms; containerised FastAPI served on Azure

LSTM trained on 48-hour ICU windows with strict patient-level train/test splits to prevent temporal leakage — a common failure mode in clinical ML. Feature engineering: forward-fill imputation, time-step delta features, rolling statistics over physiological signals.

Two complementary models: LSTM for sequence modelling, XGBoost for explainability. Tree SHAP values are faster and more interpretable than gradient attribution on recurrent networks, making XGBoost the right choice for the clinician-facing dashboard.

Introduce a feature store (Feast or Hopsworks) to decouple feature engineering from training. Add multi-task learning for additional clinical outcomes. Replace ad-hoc monitoring with a proper ML observability platform like Arize or WhyLabs.

Clinical notes are dense and unstructured — discharge summaries, progress notes, radiology reports. Keyword search fails on medical abbreviations and terminology. A semantic QA system needs to retrieve the right passages and generate grounded answers without hallucinating clinical facts.

Privacy: local LLaMA generation keeps all inference on-premise. No patient data leaves the network. Built on public datasets (PubMedQA, MedQA, synthetic clinical notes) for open reproducibility. Also used DSPy-driven prompt optimisation to generate augmented clinical notes grounded in historical data, expanding rare-case coverage in offline evaluations.

• Precision@k retrieval evaluation with configurable k and score thresholds
• Answer faithfulness scoring comparing generated answers to reference ground truth
• Hallucination rate tracked as a first-class metric — not an afterthought
• DSPy compiler automatically searches demonstration space; outperforms manual prompt templates
• Full Docker Compose stack — clone and run in one command, no proprietary data needed
• W&B experiment tracking + DVC data versioning for reproducible lineage

Bio_ClinicalBERT over general encoders: domain-adapted on MIMIC notes, it substantially outperforms sentence-transformers on clinical abbreviation resolution and entity matching. Hybrid retrieval adds keyword recall for rare drug names and ICD codes that dense search misses.

DSPy over raw prompting: the DSPy compiler searches the few-shot demonstration space programmatically. You define the signature; DSPy finds the optimal examples. This cuts prompt engineering iteration time significantly and produces more robust chains.

Local LLaMA over API: privacy-preserving, no data egress, reproducible behaviour across runs, zero inference cost at scale.

Aixigo AG is a wealth management software company. The Smart Portfolio Insider is an LLM-powered analytics layer over time-series financial dataframes, allowing analysts to query portfolio data in natural language. Schema-aware prompt templates ensure the LLM understands the financial data model. Output guardrails validate generated analysis before it reaches clients. Streaming I/O keeps the UX responsive on large portfolio datasets.

• Smart Portfolio Insider shipped to production and integrated into the Bloxx wealth platform
• ML microservices (PyTorch, scikit-learn) via Flask/FastAPI + Docker with GitLab/Jenkins CI — significantly reduced manual deployment overhead
• Shell/Python ETL pipelines for large-scale portfolio data; improved data freshness for downstream model training
• Automated testing (FitNesse), developer documentation (MkDocs), Agile delivery via Jira sprints
• Java API services for platform integration

A public demo replicates the same LLM + dataframe pattern using Yahoo Finance data — fully open-source, no proprietary data required. Demonstrates schema-aware prompting, output validation, and reproducible evaluation.

Applied the three-tier process mining methodology to an IT service desk ticket system. Descriptive analysis revealed the actual process model from event logs (vs. the assumed process). Predictive analysis identified ticket attributes correlating with SLA breaches. Prescriptive analysis recommended routing and prioritisation changes.

• Process discovery: actual vs. intended ticket resolution flow from raw event logs
• Conformance checking to flag deviations and non-compliant cases
• Bottleneck identification with queue time and waiting time breakdown per activity
• KPI impact reporting to stakeholders — actionable recommendations for process improvement