Grounded generation
Local LLM inference with retrieval, citations, provenance, and no-evidence handling.
AI / ML ENGINEERING PORTFOLIO
Building AI systems that show their work.
I build evidence-grounded AI applications and evaluation-aware machine-learning systems with Python.
Portfolio boundary · Production deployment, cloud ownership, user scale, and business impact are not claimed unless the linked repository provides evidence.
BUILD MAP evidence-led
01DOCUMENTSPDF / OCR
provenance
provenance
02RETRIEVALhybrid search
pgvector
pgvector
03MODELSRAG / fraud
evaluation
evaluation
04APPSFastAPI /
Streamlit
Streamlit
Retrieval systems
Dense and hybrid search with pgvector, rank fusion, reranking, and access controls.
Evaluation-aware ML
Leakage auditing, temporal validation, imbalance handling, threshold tuning, and task-appropriate metrics.
AI application engineering
FastAPI services, Streamlit scoring, serialized artifacts, tests, and readiness checks.
SELECTED WORK
Four systems worth opening.
Curated from the repositories with the clearest implementation and evaluation evidence.
01 · RAG / WORKFLOW PLATFORMLOCAL-FIRST
AI-Powered Port Management System
Document AI platform using PDF/OCR ingestion, PostgreSQL full-text search, pgvector, rank fusion, reranking, ACL filtering, page-level provenance, citation validation, and local Ollama generation.
DOCUMENTED CHECKPOINTAnyHit@5 0.89 · EvidenceCoverage@5 0.8510/10 mapped facts covered · 9/9 citation-valid replays
Open repository
02 · FRAUD MLDECISION SUPPORT
ClaimShield
TensorFlow/Keras insurance-fraud workflow with leakage-aware preprocessing, training-only imbalance handling, validation-based threshold selection, and reloadable inference artifacts.
TEST CHECKPOINTROC-AUC 0.8161 · PR-AUC 0.1829Fraud recall 0.8811 at threshold 0.30
Open repository
03 · TABULAR MLTIME-AWARE
Insurance Fraud & Claim Risk
Reusable local pipeline with temporal validation, leakage auditing, feature engineering, imbalance handling, threshold tuning, business-loss analysis, and Streamlit scoring.
STATUSEvaluation-aware local pipelineRecorded test-window checkpoint is weak; no production success is claimed.
Open repository
04 · DATA / ML ENGINEERINGCASE STUDY
Smart Education Analytics
OULAD-based workflow covering RDDs, DataFrames, Spark SQL, ETL feature preparation, Spark ML, Docker packaging, Kubernetes manifests, and CI checks.
HOLDOUT CHECKPOINTAUC 0.9706 · F1 0.9142Accuracy 0.9142 for the provided case-study outcome proxy
Open repository
Also see HealthBot for structured symptom matching and dense NLP retrieval. It is an educational local demo, not a diagnosis system.
ENGINEERING APPROACH
Make the system inspectable.
01
Ground the output
Connect generated answers to retrieved context, citations, provenance, and explicit no-evidence behavior.
02
Measure the boundary
Keep evaluation splits, thresholds, artifacts, and limitations visible beside the result.
03
Ship responsibly
Use tests, configuration, readiness checks, and clear boundaries around local or unverified deployment.
LET'S CONNECT
Interested in practical AI systems?
Open to AI Engineer, ML Engineer, GenAI Engineer, LLM Engineer, and Software Engineer opportunities.