Available for select advisory

Hi, I’m Imran.

I build applied AI/ML systems. Currently co-founding VEZRAN — shipping Zyberpol, an agentic Security Operations Center platform. Previously Senior Data Scientist at Starbucks and FedEx.

I write about production ML, contribute to open-source AI infrastructure, and disclose security research to the projects I depend on.

Read the writing → See 26+ merged PRs

26+
merged OSS PRs: 8+
years shipping ML in production: 1
agentic SOC platform, in production: 3
companies (incl. one I co-founded)

Writing — long-form essays

all →

Featured

Evaluating multi-agent systems when ground truth is incomplete

Three eval modes — output, trajectory, and side effect — and the practical pattern we use when there's no single right answer.

May 25, 2026 · 4 min read

agents eval production-ml

Open source — 26 merged PRs, 4 open

all →

dowhy

Determinism + significance-level API

FLAML

Reproducibility audit — 6 PRs, zero review comments

MAPIE

.reset() trilogy + optimize_beta bug fix

Shapash

SmartPredictor production-hardening series

PyRIT

Scorer + orchestration fixes

imbalanced-learn

SMOTENC encoder + sample_weight

Security research

Coordinated disclosure 2026-06-30

HF Transformers path-traversal cluster

Four-site path-traversal cluster in checkpoint loading — extends the #46097 fix. Disclosed via Huntr + security@huggingface.co with proof-of-concept and remediation.

Huntr #38583 · HuggingFace

Security benchmark 2026-06

DSPy Security Bench

Reproducible security benchmark for DSPy pipelines — prompt injection, jailbreak, and adversarial-example attacks across representative DSPy programs.

immu4989/dspy-security-bench · Open source (own repo)