Mukund Agarwal

An ongoing side research project on drug interaction: given a molecular structure, predict how a compound is absorbed and cleared, and how two drugs taken together change each other's behaviour. Machine-learned property models feed a simulation of the body over time, so a prediction comes with a mechanism rather than a label.

Benchmark v1.2.0: 79.4% of predictions land within 2-fold of published values, and roughly 80% within 2× across 30 clinically studied drug pairs. The harness prints its own failures — theophylline, atorvastatin, metoprolol. Evaluation design reviewed with Dr. Joga Gobburu, former Director of the FDA Division of Pharmacometrics.

A typed seven-agent research runtime that argues with itself: tenant-scoped retrieval, adversarial hypothesis gates, claim-level grounding. High-impact unsupported claims are redacted, and hypotheses must survive a debate gate before they ship. The full verification path runs offline — no API keys, no database.

A zero-trust policy engine for MCP tool calls: default-deny TOML policy, filesystem root confinement, private-address blocking, prompt-injection detection in nested arguments, credential redaction — and a SHA-256 hash-chained, tamper-evident audit log of everything an agent tried to do.

Break your agent on purpose. Deterministic fault injection for tool-using agents — timeouts, provider failures, malformed output, stale memory, repeated calls — with content-addressed run IDs and hash-chained traces, comparing baseline against resilient strategies on pass rate and recovery.

Post-training you can reproduce: a tested SFT/DPO/GRPO harness with deterministic splits, SHA-256 dataset fingerprints, and promotion gates. In one controlled Tesla T4 run, gradient checkpointing cut peak VRAM 21.1% while held-out loss stayed within 0.0025, at a 26.1% throughput tradeoff.

A city of seeded Q-learning agents under audit: prioritized replay, reward shaping, resource and labor markets. Counterfactual tests prove identity metadata never alters aggregates, and generated analyses cannot approve themselves out of the human-review gate.