Ebin's Profound AI Rep
Ebin turns model claims and actions into systems that can be tested, traced, and audited.
Ebin Babu Thomas
Edges
Designing Verifiable AI Systems
Ebin designs AI systems whose evidence and actions can be checked by deterministic code outside the model. In ProofPack, hashes and verbatim quote checks govern admissible evidence. In IncidentGate, persistent state, traces, policy gates, and single-use approvals make agent behavior inspectable. The recurring decision is to separate model judgment from the controls that establish whether a claim or action is valid.
Spotted in 2 Stories
Designing Failure-Oriented Evaluations
Ebin defines what should fail, freezes the criteria, and builds tests that force the system through those conditions. IncidentGate tested interruption and composed attacks. diffing-agent-bench planted model differences behind a sealed answer key, while ProofPack simulated fabricated evidence. His evaluations produce decision-grade findings, including null and negative results that remain part of the published record.
Spotted in 3 Stories
Scoping Under Technical Ambiguity
Ebin converts uncertain briefs into fixed acceptance criteria, small experiments, and a clear product decision. For the HACCP conversion feature, he tested several representations and identified the fine-tuning requirement. Across agency MVPs, he used short proofs of concept and frequent client feedback. This approach keeps the objective stable while allowing the implementation to change as evidence arrives.
Shipping Applied AI Backends
Ebin builds Python and FastAPI services, retrieval flows, model integrations, and cloud deployments. He has shipped NL-to-SQL search, Qdrant document retrieval, speech assessment on AWS Fargate, and a vLLM prototype on GKE. His independent work adds LangGraph, MCP, Playwright, PostgreSQL, and model fine-tuning. The range supports end-to-end implementation from a small proof through a deployed backend or reproducible evaluation harness.
Spotted in 1 Story