Evidence-led case studies covering agentic systems, applied AI research and responsible production engineering, with implementation roles and evaluation limits made explicit.
Extended FinRL–DeepSeek with a confidence-weighted, turbulence-aware signal layer for constrained PPO trading and evaluated the system through deterministic and 10-seed stochastic backtests.
Implementation context and limits
Individual Phase 1 extension and sole report authorship, with explicit attribution to the upstream FinRL–DeepSeek project.
Results are historical backtests, not live-trading performance or investment advice; upstream attribution and documented evaluation limitations must remain visible.
Ahmed’s role
Individual Phase 1 extension and sole report authorship, with explicit attribution to the upstream FinRL–DeepSeek project.
Work status
Individual work by Ahmed
Verified evidence
Deterministic EP90 annual return: 17.12%
Method: Deterministic EP90 historical backtest.
Ten-seed stochastic mean annual return: 16.843% ± 3.517%
Method: Ten-seed stochastic historical backtest; mean ± standard deviation.
Ten-seed stochastic mean maximum drawdown: −25.921% ± 1.671%
Method: Ten-seed stochastic historical backtest; mean ± standard deviation.
Deterministic total return: 119.86%
Method: Deterministic EP90 historical backtest.
Deterministic Sharpe ratio: 0.750
Method: Deterministic EP90 historical backtest.
Ten-seed stochastic mean Sharpe ratio: 0.878 ± 0.173
Method: Ten-seed stochastic historical backtest; mean ± standard deviation.
A multi-agent Socratic tutoring system with persistent learner state, learning history, adaptive guidance and user-triggered fairness-audit functionality.
Implementation context and limits
Verified implementation includes a planner agent, Socratic tutor, learning-signal extractor, fairness reviewer, controller and state transitions, SQLite persistence, a CLI and a Streamlit interface.
Fairness audits are user-triggered; no learning-effectiveness or fairness-outcome evaluation has been completed.
Ahmed’s role
Individual academic project.
Work status
Individual work by Ahmed
Verified evidence
Multi-agent architecture: Planner agent; Socratic tutor; learning-signal extractor; fairness reviewer; controller and state transitions.
Method: Implementation-scope review.
Persistent learner state: SQLite persistence supports learner state and learning history across sessions.
Method: Implementation-scope review of the CLI and Streamlit interfaces.
Responsible-AI boundary: Fairness audits are user-triggered; learning effectiveness and fairness outcomes remain not evaluated.
Method: Implementation-scope review; no learning-effectiveness or fairness-outcome result is claimed.