Deepam Ahuja

I'm a backend engineer working on distributed systems, LLM agent architectures, and AI safety research.

At Goldman Sachs, I worked on a post-execution reconciliation service matching internal trade records against confirmations from a third-party confirmation venue, around 56,000 trades a week, several XML lifecycle messages each, on an event-driven architecture over Kafka. Java and Spring Boot. Most of what I do now is agentic systems and how they fail. ChainFlow is a multi-agent procurement platform that triggers RFQs, routes approvals and dispatches purchase orders on its own with decision authority scoped per role in the orchestration layer, so agents act independently but not outside their permission boundary, and every action is auditable. FairLens takes the same view of enforcement: it audits models against defined thresholds and returns a pass/fail exit code, so a breach blocks the deployment instead of waiting on someone to run a review.

On the research side I work on mechanistic interpretability of trust in language models - linear probes, activation patching, sparse autoencoders. I've also built a peer-to-peer safety monitoring mesh for detecting cross-session fragmentation attacks, and an eval harness that probes whether an agent behaves differently when it can tell it's being watched.

Elsewhere: I reconciled 2.4 million records with broken metadata into a usable dataset and ran cascade-risk simulation over it in PySpark, and I contribute to sktime.