Senior AI Engineer · GenAI & LLM Systems · Open-Source Contributor

"Build it end-to-end. Ship it. Then explain why it works."

Fragile Safety: Automated Circuit Discovery is Vulnerable to Dormant Feature Bundling Chandrashekar DP — Zenodo Preprint, June 2026

Automated circuit discovery tools used for AI safety verification are structurally vulnerable to adversarial input manipulation. A linear probe achieving 97.5% accuracy on clean data fails completely on adversarial inputs (0% detection rate). The adversarial distribution is geometrically inseparable from clean positives at the anchor token (cosine similarity 0.989). Reproduced on Pythia-410m. Practical mitigation: context-aware probing at the last sequence token achieves 100% adversarial detection with clean accuracy preserved. All experiments use TransformerLens and run on CPU.

📄 Preprint — Zenodo · 💻 Experiments code · arXiv submission pending endorsement

• TransformerLens — Contributing unit tests for architecture adapters in Google DeepMind's mechanistic interpretability library (3,500+ ⭐). 5 PRs submitted, 4 merged — covering GPT-Neo, GPT-NeoX, Apple OpenELM, LLaVA-OneVision, and LLaMA. Each PR adds production-quality test coverage (147 tests, 1,286 lines) for AI architectures used by safety researchers at Anthropic, Google DeepMind, and universities worldwide.

• enterprise-rag-patterns — Production RAG architectures from my "GenAI in Production" newsletter. Code derived from real enterprise systems — not demo code.

  • article-01: Intelligent RCA Agent with Claude API + Databricks (80% triage reduction)
  • article-02: RAG Evaluation at scale — proxy metrics, embedding drift detection, labeled eval with Claude as judge

• ComplianceShield — AI compliance assistant with PII detection, session audit logging, multi-provider LLM support, and HITL review workflows. FastAPI + Streamlit + Docker. Built to solve the trust problem in enterprise AI deployments.

• Working through nanoGPT → micrograd → minbpe (Karpathy's series) to understand transformers from raw math, not API calls
• Studying mechanistic interpretability via TransformerLens contributions — learning how attention heads and MLP circuits encode knowledge

• The LLM observability gap: LLMs are in production everywhere. Engineers have nothing equivalent to what researchers have in TransformerLens. No production-grade "Sentry for model reasoning." That's the problem I keep coming back to.

• Why evaluation is harder than training: Writing the RAG evaluation framework taught me that measuring whether an LLM is correct is a deeper problem than making it correct. Most teams skip it. The ones who don't, ship reliable AI.

• Interpretability as infrastructure: TransformerLens contributions are clarifying something — the people who build the measurement tools shape what the whole field builds next. Open-source research infrastructure is underrated leverage.

• Foundations vs. APIs: There's a big difference between engineers who use LLMs and engineers who understand them. Working through micrograd → nanoGPT is making that difference concrete for me.

AI & Research

• Concrete Problems in AI Safety — Amodei et al.
• Attention Is All You Need — Vaswani et al.
• Anthropic's Mechanistic Interpretability papers (superposition, features, circuits)
• Neel Nanda's mech interp blog posts

Engineering

• Designing Machine Learning Systems — Chip Huyen
• Building production RAG systems (hands-on, via enterprise-rag-patterns)

Just for interest

• Thinking, Fast and Slow — Daniel Kahneman
• My wife's opinions on everything (ongoing study, steep difficulty)

⚡ Fun fact: I studied both AI interpretability research infrastructure and enterprise root cause analysis in the same week — and they're the same problem.