Abdoul27/phase3-track-a-telco-troubleshooting

Phase 3 Track A — Telco Troubleshooting Agentic Challenge

Outer Banks 5 (solo team) — ITU × Zindi 2026

Files in this repository

File	Purpose
code.zip	The submission archive — extract and follow README.md inside
TECHNICAL_REPORT.pdf	Technical report describing the agentic workflow

What's inside code.zip

code.zip  (37 KB, 12 files)
├── README.md           reproduction instructions
├── run.sh              entry shell wrapper
├── run.py              driver — iterates scenarios, writes result/
├── requirements.txt    4 pip packages
├── models/
│   └── deploy.sh       vLLM startup for Qwen3.5-35B-A3B
└── src/
    ├── runtime_logger.py   verbatim from spec
    ├── phase2_pipeline.py  16-rule deterministic solver
    ├── sandbox.py          22-endpoint HTTPS client
    ├── tools.py            15 tools + analytic helpers
    ├── prompts.py          unified agent prompt
    └── agent.py            Qwen tool-calling loop with safety nets

Architecture summary

Qwen3.5-35B-A3B is the orchestrating brain; fifteen specialised tools are its hands. The brain calls recommend_action() (the solver) first on every scenario, reads the routing flag, and selects one of three investigation protocols: accept the solver's recommendation directly, refine a binary geometric choice via select_azimuth(), or freely explore the 24-option remediation menu via a fault-specific analyse_* tool.

Six layers of defence (streamed outputs, bounded loop, empty retry, tier-aware heuristic fallback, truncation safety net, last-ditch direct solver call) ensure every scenario produces a valid non-empty answer in well under the 5-minute time-discount cap.

See TECHNICAL_REPORT.pdf for the full architecture, tool catalogue, and reliability design.

How to run (summary)

pip install -r requirements.txt
$EDITOR models/deploy.sh        # set MODEL_PATH to Qwen3.5-35B-A3B weights
bash models/deploy.sh           # wait for "Application startup complete"
bash run.sh --input test_phase3.json --output result

Outputs land in result/: results.csv, traces.json, runtime.json.