Mihai Cosma

+1 (613) 413-0204 · mcosma@gmail.com · mihaicosma.com · github.com/wakamex

Senior finance and quantitative practitioner with 12 years in Canada's foreign-reserve program, including co-leading a $20 billion non-USD credit portfolio and managing credit allocation in a €14 billion portfolio. Built Python trading, fuzzing, and protocol-risk systems in DeFi. Since February 2026, published 10 applied autoresearch field reports across search, ML/RL, simulation, inference, and software automation that explicitly label evidence boundaries, failures, and promotion authority.

Python · SQL · PostgreSQL · JAX · DuckDB · agent evaluation · reinforcement learning · financial markets

Selected Applied AI and Software Work

• Autoresearch field reports Published 10 applied campaigns from February through August 2026 across search, ML/RL, simulation, inference, and software automation. The reports include negative results and distinguish adaptive discovery from promotion evidence.
• Agent Orchestration Process Built and published a Python CLI for bounded unattended work across eight coding-agent harnesses, with isolated worktrees, deadlines, least-authority profiles, exact session resume, declared artifact validation, and durable evidence.
• Tactics.md Tested 102 CPU-pipeline interventions and uncovered eight correctness defects, then evaluated 435 neural-model result bundles. Under an equal move budget, the model with the best validation loss played worst among matched candidates.
• FreeTranscribe.org Built and operate an AI transcription service with a production-priority GPU queue. A sealed promotion doubled warm inference throughput on a shared RTX 3090 without increasing errors across 295 cases.
• FlySim Built a public flight-control stack with a CPU reference environment and batched JAX reinforcement learning. Exact replay localized a simulation-to-live coverage gap; retraining eliminated 14% control clipping without weakening live safety limits.

Work Experience

Education