Hi, I'm Laela Zorana 👋
AI evaluation specialist and MLOps engineer building production-quality tools for AI training data QC, model evaluation, and performance engineering.
📍 Burr Ridge, IL, USA | 🔗 github.com/LaelaZorana | 📊 kaggle.com/laelazorana
⭐ Featured live demos
Live, deployed, and — where the model is the point — running the real model, not a stub.
| Demo | What it proves |
|---|---|
| 🎭 Emotion Spectrum API · model | Fine-tuned DistilBERT — accuracy 0.920 / macro F1 0.874 (held-out), with a confusion matrix — served by a production API (batching, Prometheus/Grafana). The demo runs the real model. |
| ⚖️ ParaPilot | Grounded, cited legal-Q&A RAG with an anti-hallucination eval — 3.8% hallucination / 85.3% groundedness on real Haiku-4.5. |
| 🛟 SupportCopilot | E-commerce support agent — 80% ticket deflection on the seeded eval. |
| 🎯 InterviewCoach | Job-description-aware interview prep with rubric-scored feedback. |
| 🛡️ Solidity Static Auditor | Rule-based SWC-mapped detectors + optional LLM remediation. |
| 🔐 LLM Security Scanner | Injection / jailbreak / PII probes + a governance report. |
| 📊 CommercePipeline | DuckDB warehouse with data-quality gates that fail closed. |
🛠️ What I Build
AI Training Data Quality Control
Tools for the full AI training data QC pipeline — scenario validation, RLHF preference rating, content moderation, UI annotation, and SCORM package validation.
| Repo | What it does |
|---|---|
| ai-agent-scenario-qc | Reviews AI agent training scenarios (JSON) against an 8-criterion weighted rubric. Detects undeclared tools, undefined personas, token bloat. |
| rlhf-pairwise-rater | Pairwise rating CLI on 4 axes (helpfulness, harmlessness, accuracy, instruction-following) with self-consistency check and Cohen's kappa. |
| content-policy-rater | Rubric-based content moderation reviewer with per-criterion inter-rater agreement. |
| ui-annotation-toolkit | Screenshot annotation with forced observation/inference separation. |
| scorm-qa-validator | Validates SCORM e-learning packages locally before LMS upload. |
MLOps & Performance Engineering
| Repo | What it does |
|---|---|
| mlops-training-pipeline | Composable MLOps toolkit: training loops, experiment tracking, HuggingFace evaluation, JAX trainer, DDP/FSDP config. |
| triton-kernel-lab | Custom GPU kernels in OpenAI Triton — fused softmax, GELU, tiled matmul — with PyTorch benchmarks. |
| python-perf-optimizer | Python performance profiling, latency/throughput benchmarking, memory optimization, and severity-tagged optimization advisor. |
🎓 Background
- MS Cloud Computing — Purdue University (In Progress, 2026)
- AI Agents and Agentic AI Architecture in Python — Vanderbilt University (Dec 2025)
- Lean Six Sigma — Kennesaw State University (Jan 2026)
- 10+ years in content quality, compliance auditing, and technical QA at Deutsche Bank, Morningstar, Tata Capital, Altisource, and Accenture
🤝 Connect
- GitHub: github.com/LaelaZorana
- DagsHub: dagshub.com/laelazorana
- LinkedIn: linkedin.com/in/laela-zorana-362309114
- Kaggle: kaggle.com/laelazorana
- Email: zoranalaela9@gmail.com
Who built this
I am Laela Zorana. I write GPU kernels and measure them on real hardware, on AMD Instinct cards and on AWS Trainium, and most of my work sits at the point where two correct answers disagree and something has to decide which one to trust.
If this was useful, a like puts it in front of the next person looking for the same thing, and following means you see what I measure next. If you are using it for something, I would like to hear what.
Where the rest of my work lives.
- Hugging Face, https://huggingface.co/LaelaZorana
- GitHub, https://github.com/LaelaZorana
- Website, https://codezorana.ai
- LinkedIn, https://www.linkedin.com/in/laela-zorana-362309114