Portfolio

Advanced ML & systems implementation

These are selective, full-stack implementation examples—real-time serving, batch scoring, retrieval, feature alignment, and batch monitoring—meant to sit above the primary storefront (prep, KPIs, reporting automation). They show how milestones are bounded, handed over, and run again with the same contract—not generic demos or coursework.

Key architecture notes (show)

Explicit batch versus API boundaries: repeatable jobs versus request-time inference, with artefacts and versioning you can pin
From raw extract to scored output: preprocessing locked to fit-to-score conventions and failure behaviour you can script
Retrieval-assisted Q&A patterns with citations and guardrails—as a reference architecture, gated for real client data and scope elsewhere
Operational maturity touches: drift and quality batches, structured artefacts, rerun-friendly reports—not magic dashboards

VahdetLabs sells scoped engagements first on practical data workflows; this page collects deeper ML and systems implementations for buyers who already need—or will soon need—premium milestones. Repos include tests and deploy notes; contracted work stays defined in writing on Services.

GitHub repositories — supplementary context; cards open expanded notes on Work and still expose repository and demo in each footer.

Projects in this track

Customer churn prediction system

Shows a production-style pattern: one pinned model artefact, fixed probability and tier outputs, and an HTTP surface your stack can call—without silent notebook drift on the live path.

FastAPI
scikit-learn
Classification
Joblib

Per-account churn probability, tier, and routing flag are served from a single frozen artefact and agreed cutoffs. Live inference never depends on ad-hoc notebook runs; training and ad-hoc exploration stay off the request path by design.

Public deploy reflects an end-to-end reference stack (model package, API, TLS, health checks). The bundled dataset illustrates the architecture only—your engagement would use your data, written acceptance checks, and separate milestones for retrain, drift review, or always-on observability.

GitHubLive demo

Batch scoring pipeline

Scores many records in one run on a schedule, with the same rules each time instead of one-off checks.

Python
Batch
pandas
pytest
joblib

Batch job scores CSV rows with training-aligned preprocessing and writes score, label, model version, and timestamp on each row. Built for scheduled runs with deterministic row output and explicit failure exits.

The live URL illustrates static input/output conventions; scoring runs locally or inside your orchestration. Optional manifest, non-zero exit on failure, and a documented row audit path verify runs end to end.

GitHubLive demo

RAG document intelligence QA

Answers questions from your own documents in plain language and points to the exact source for each answer.

FastAPI
FAISS
sentence-transformers
Docker
OpenAI / Ollama

Ingest chunks documents for retrieval, then answers questions using retrieved context and a hosted model. Each reply cites document, page, and chunk identifiers so answers stay traceable.

The deployed API demonstrates rate limits, health checks, optional keys, retrieval floors, and CORS—the sort of operational scaffolding a scoped retrieval pilot expects. Bounded scope, corpus hygiene, and client-side ownership remain contract questions, not turnkey rentals.

GitHubLive demo

Train–score feature compilation pipeline

Keeps training and downstream scoring aligned on the same engineered fields so exported rows match what the model was fitted on.

pandas
pytest
FastAPI
Batch pipeline

A batch compilation step turns approved raw extracts into a versioned feature table under locked column contracts. Fit and score both load the same built artefact so transforms cannot silently diverge across environments.

Useful when you treat reliability as engineering: rejects bad batches early, validates schema and duplicates, and can expose a thin HTTP helper for inspecting transforms—not a sprawling vendor feature platform. Fits buyers who already buy into batch-first ML milestones.

GitHubLive demo

ML monitoring & batch data quality

Shows when freshly scored batches or inputs drift away from an agreed baseline so teams react before KPIs silently rot.

pandas
pytest
Streamlit
PSI / KS
JSON schema

Compares each new batch against a frozen reference distribution and emits structured summaries for inputs, categorical shifts, and score behaviour.

The browser view is only a viewer; the substantive output is repeatable batch artefacts (reports, drift JSON). Scope stays anchored to batched files you control—ideal as a maturity add-on alongside batch scoring engagements, not a generic live APM substitute.

GitHubLive demo

Each card lists its GitHub repository and live demo on a vahdetlabs.com subdomain. Repos remain the source-of-truth if a deploy URL ever changes.