anupamagarwal001/amc_allocator_env

1# Round 2 Demo Flow2 3This is the practical live-demo sequence to use on campus.4 5## Goal6 7Show four things clearly:8 91. this is a real multi-agent environment102. reward is objective and hard to game113. the Portfolio Manager is trainable124. the training result is measurable13 14## Demo Length15 16Target: `3 minutes`17 18## Screen Order19 20### Screen 1: Problem framing21 22Show:23 24- repo README or spec summary25 26Say:27 28"Round 1 was a single-agent allocator. For Round 2 we turned that into a multi-agent investment committee where the Portfolio Manager learns by coordinating with Research and Risk."29 30### Screen 2: Environment structure31 32Show:33 34- the task list35- PM actions36- reward breakdown37 38Say:39 40"The PM can query Research, query Risk, allocate, hold, or move to cash. We score not only returns, but also compliance, risk response, and information usage."41 42### Screen 3: Baselines43 44Show:45 46- `baseline_report.json` or terminal table47 48Say:49 50"Before training, the heuristic policy scores `0.4084` overall while random is `0.2504`. So the environment already has a meaningful difficulty and signal structure."51 52### Screen 4: Training proof53 54Show:55 56- `grpo_behavior_sample.svg`57- `reward_curve.png`58- `judging_report.md`59 60Say:61 62"We ran a verified T4 LoRA smoke training job with Qwen3-0.6B. The reward started at `0.0193` and ended at `0.0275`, with the best step at step `4`. The preserved completion trace shows why: the verifier gives higher reward when the PM gathers both Risk and Research before taking exposure."63 64### Screen 5: Why it matters65 66Show:67 68- `onsite_demo_summary.md`69 70Say:71 72"This environment is useful because it teaches an agent to act in a structured professional workflow, not just to predict one-step allocations."73 74## If Judges Ask For More Detail75 76Open:77 78- [committee_grpo_train.py](/Users/anuagar/Desktop/dev/amc_allocator_env/training/committee_grpo_train.py)79- [committee_artifacts.py](/Users/anuagar/Desktop/dev/amc_allocator_env/training/committee_artifacts.py)80- [server/amc_environment.py](/Users/anuagar/Desktop/dev/amc_allocator_env/server/amc_environment.py)81 82Focus on:83 84- only the PM is trainable85- Research and Risk are scripted environment actors86- reward uses multiple verifier-style components87- training outputs are exported automatically into judge-facing artifacts88 89## What Not To Waste Time On90 91Do not spend live demo time on:92 93- full code walkthroughs94- raw notebook debugging95- long theoretical RL explanations96- long finance explanations97 98Keep it on:99 100- world design101- reward design102- measured improvement103- why this is aligned with the hackathon themes104 105## Fallback Demo106 107If compute fails onsite:108 1091. show the environment and task structure1102. show heuristic vs random baselines1113. show the previously verified T4 smoke result1124. explain the exact training config used113 114That is still a coherent story.115