Wasim Akram
Principal Data & AI Engineer @ VML MAP (WPP) · P1 Affiliated Researcher @ AI Centre Denmark · Copenhagen 🇩🇰
I build production-grade agentic AI systems and resilient data platforms while studying how sparsity and modularity emerge inside dense language models—and how to turn that structure into measurable efficiency gains.
Snapshot
- Day role: Enterprise ML/AI platforms, cost-aware inference, activation-aware monitoring.
- Research track: Scaling Specialization in Dense LMs—activation sparsity, sparse autoencoders, dynamic-k execution.
- Study plan: Working through Antikythera—What Is Intelligence? and replicating Anthropic interpretability studies to stay grounded before extending the frontier.
- Interests: Mechanistic interpretability, sparse autoencoders, scaling laws, activation-efficient inference, trustworthy evaluation.
Links
Current Research
- Measure specialization: Activation sparsity, specialization index, and modularity metrics across the Pythia scaling suite (70M→1B+).
- Recover interpretable features: Train sparse autoencoders, score monosemanticity, and map circuits tied to IOI, arithmetic, induction, POS/NER probes.
- Exploit structure: Prototype dynamic-k gating, validate causal roles via targeted ablations, and report throughput vs perplexity trade-offs.
Stay hungry. Stay foolish.