@alexey-safonov-liminal
Independent AI safety researcher building deterministic oversight, trace evaluation, and causal memory tools for agent systems.
https://github.com/safal207$0 in pending offers
I work on AI safety infrastructure for agent systems: deterministic trace protocols, causal memory layers, adversarial evaluation, and failure localization. Before pivoting into AI safety, I spent 12+ years in fintech QA, where I worked on high-reliability testing, auditability, and production-grade failure analysis. My goal is to turn vague concerns about agent misalignment into concrete, reusable open-source artifacts: benchmarks, trace tooling, and oversight libraries.
pending admin approval
Alex Safonov
3 days ago
90-day execution plan
Days 1–30: finalize trace schema, integrate one framework, define failure taxonomy.
Days 31–60: build initial adversarial corpus, implement CML graph logic, run first baseline comparisons.
Days 61–90: expand benchmark, publish evaluation write-up, release reusable library + documentation.
Alex Safonov
3 days ago
Update:
I have active related applications under evaluation with LTFF, Open Philanthropy / Coefficient-affiliated funding, and NLNet, and I am currently clarifying fiscal sponsorship structure for an external SFF application. Current focus is tightening benchmark scope, deliverables, and release plan for LTP + CML.
Alex Safonov
3 days ago
Update:
I have active related applications under evaluation with LTFF, Open Philanthropy / Coefficient-affiliated funding, and NLNet, and I am currently clarifying fiscal sponsorship structure for an external SFF application. Current focus is tightening benchmark scope, deliverables, and release plan for LTP + CML.
Alex Safonov
3 days ago
Update: We are aligning fiscal sponsorship for an external SFF application. This project is gaining multi-fund interest.