Migration Cutover Decision Advisor — Powered by IBM Bob + IBM watsonx.ai Granite.
🚀 Live demo: bob-the-judge-production.up.railway.app
Built for the IBM Bob Hackathon 2026 (Lablab.ai, May 15–17).
Bob writes the code. Bob ships it. Bob the Judge tells Bob when to flip the switch.
End-to-end IBM stack: IBM Bob as the development partner (5 task sessions in bob_sessions/), IBM Granite (ibm/granite-4-h-small) running live on watsonx.ai for the runtime inference. Every Plan / Code / Ask / Orchestrator response in the dashboard is a real Granite call.
Fresh dashboard. Topbar shows ● LIVE Bob | ● watsonx.ai. Sidebar lists IBM Bob API + watsonx.ai (Granite) + MCP Server, plus the Tenant Profile selector (Tier1 Global = 99% threshold, $0.01 tolerance, 1.5σ anomaly).
Run Analysis with Demo mode on the Tier 1 Global profile. 80 transactions through legacy + modern. 2 functions blocked (International Wire Transfer, High-Value Wire Transfer), $4.8M exposure, 1 anomaly (>2σ). Bob the Judge says HOLD — DO NOT CUT.
Click Apply Bob's Patch & Re-Analyse. Bob deploys the parallel-run compatibility patch (modern fees align with legacy during dual-run window). Re-runs analysis. All 4 functions cleared, zero exposure, READY FOR CUTOVER.
Bob explains the patch: real-time call to ibm/granite-4-h-small on watsonx.ai. Session ID WATSONX-... for full traceability. Granite generates a compatibility patch design citing FFIEC, Basel III, PSD2 by name.
Enterprise migrations from COBOL to modern stacks stall in dual-run for months because no tool answers the question: when is it safe to cut over?
Bob the Judge produces a per-function readiness verdict (SAFE TO CUT / DO NOT CUT) by routing live traffic through both legacy and modern systems, scoring functional parity with a Wilson 95% confidence interval, flagging anomalies (>2σ), and exporting a regulator-grade audit PDF.
It runs Bob in all four modes — Plan, Code, Ask, Orchestrator — and exposes itself as an MCP server so Bob can query it directly from the IDE.
| Feature | Description |
|---|---|
| Per-function verdicts | SAFE TO CUT / DO NOT CUT per payment function — not a single system score |
| Wilson 95% CI + Bootstrap + Bayesian | Three statistical methods cross-validate readiness (see parity/confidence.py) |
| >2σ anomaly flagging | Fee divergences beyond 2 standard deviations flagged before production |
| Per-tenant calibration | 4 reference profiles (tier1 / tier2 / community / demo) with FFIEC, Basel III, PSD2 alignment — switch in the sidebar |
| Drift detection (CUSUM + EWMA) | SPC charts on run history catch sustained parity degradation across the migration window |
| Regulatory compliance framework | FFIEC IT Handbook, Basel III, PSD2, SWIFT gpi, SOX 404 mapping with pass/fail scoring |
| Live IBM Granite inference | Plan / Code / Ask / Orchestrator responses generated by ibm/granite-4-h-small on watsonx.ai |
| Parity Radar Chart | Spider chart comparing each function's parity rate against the cut threshold |
| IBM Carbon scanning animation | Terminal-style 4-step progress indicator during analysis |
| Bob Session ID in PDF | Full chain-of-custody: Bug → Bob Fix → Re-test → Signed off. Regulator-grade. |
| One-click audit PDF | Executive summary, tenant risk profile, divergence log, Bob Session ID, sign-off page |
| MCP server | Bob queries Bob the Judge directly from the IDE — 6 tools exposed (Bob + tenant introspection) |
Traffic Generator → Legacy Bank ┐
├→ Parity Engine → Scoring → Verdict
→ Modern Bank ┘ ↓
Bob (Plan/Code/Ask/Orchestrator)
↓
Audit PDF + MCP Server
| Component | File | Description |
|---|---|---|
| Legacy Bank | services/legacy_bank.py |
FastAPI port 8001, COBOL-style fee logic |
| Modern Bank | services/modern_bank.py |
FastAPI port 8002, modern Python, /admin/apply_patch |
| Parity Engine | parity/parity_engine.py |
N transactions through both, diffs, >2σ anomaly flagging |
| Scoring | parity/scoring.py |
Wilson 95% CI on parity rate per function |
| Bob Client | bob/client.py |
IBM Granite via watsonx.ai (primary) → Bob REST API (fallback) → intelligent mock (last resort) |
| Tenant Config | parity/tenant_config.py + config/tenants.json |
4 reference bank tier profiles with regulatory rationale |
| Confidence Methods | parity/confidence.py |
Wilson, Bootstrap, Bayesian credible intervals, funnel plot |
| Drift Detection | parity/drift_detection.py |
SQLite-backed run history + CUSUM + EWMA SPC |
| Compliance | parity/compliance.py |
FFIEC / Basel III / PSD2 / SWIFT gpi / SOX 404 mapping |
| Dashboard | dashboard.py |
Streamlit, IBM Carbon Design System, port 8501 |
| MCP Server | mcp_server.py |
FastMCP, 4 tools for Bob IDE |
| Audit PDF | audit/pdf_report.py |
ReportLab, regulator-grade sign-off document |
Option A — One double-click (Windows)
BobTheJudge.exe
Clears ports 8001/8002/8501 → starts both bank services → opens dashboard in browser.
Requires Python 3.11+ in PATH. The exe is a launcher, not a self-contained bundle. If you see a "Python not found" error, install Python 3.11+ and ensure it is added to PATH.
Option B — Python
# Install deps
pip install -r requirements.txt
# Launch everything
python launch.pyEnabling live IBM Granite inference (optional but recommended):
# Copy .env.example to .env and add your watsonx.ai credentials
WATSONX_API_KEY=...
WATSONX_PROJECT_ID=...
WATSONX_URL=https://us-south.ml.cloud.ibm.comWithout these, the Bob client transparently falls back to the bundled intelligent mock — every demo path still works.
Once the dashboard is up at http://localhost:8501:
- Click Run Analysis — 80 transactions through both systems
- Review the 4 scorecards (2 SAFE, 2 DO NOT CUT by design for demo clarity)
- Click Apply Bob's Patch & Re-Analyse — all 4 flip GREEN (the wow moment)
- Click Export Audit PDF — regulator sign-off document generated instantly
Bob can query Bob the Judge directly using bob-mcp-config.json. Four tools exposed:
| Tool | Purpose |
|---|---|
analyze_cutover_readiness(n) |
Run a fresh analysis, return full report |
get_function_verdict(name) |
Per-function root cause + recommended action |
get_risk_summary() |
Executive dollar-exposure briefing |
list_monitored_functions() |
Lists all 4 payment functions |
python mcp_server.pyBob the Judge extends watsonx.governance Q1 2026 Agent Monitoring & Insights into the COBOL → Java cutover window:
| watsonx.governance | Bob the Judge |
|---|---|
| Real-time agent decision tracking | Per-function readiness verdicts on live transactions |
| Threshold breach alerts | >2σ anomaly flagging + Wilson 95% CI gate |
| Continuous compliance reporting | One-click regulator-grade PDF |
Designed to plug into watsonx Orchestrate workflows.
The bob_sessions/ folder contains 5 IBM Bob task session reports captured on the hackathon-provisioned account (ibm-coding-challenge-xxx), each with the exported markdown plus a screenshot of the task session consumption summary, per [IBM Bob Hackathon Guide page 18]:
| Session | Bob Mode | Outcome |
|---|---|---|
A_plan.md |
Plan / Agent | Roadmap + Phase 1 (per-tenant calibration) implemented end-to-end by Bob — created parity/tenant_config.py, config/tenants.json, integrated into engine + scoring |
B_code.md |
Code / Edit | Bob added format_score_summary() helper to parity/scoring.py for CLI/log output |
C_ask.md |
Ask | Regulator-facing memo on per-function vs system-level cutover decisions, citing FFIEC, Basel III, PSD2, SWIFT gpi, NIST AI RMF |
D_orchestrato_sum.md + sub-tasks |
Orchestrator | 4-stage cutover pipeline (ANALYSER → RISK-SCOUT → FIX-GEN → REPORTER) with Bob delegating each stage to specialised sub-tasks |
E_review.md |
Ask (code review) | Bob audited the entire enhanced codebase and produced a prioritised findings report (CRITICAL / HIGH / MEDIUM / LOW) |
Bob also wrote the Phase 2-4 modules (parity/confidence.py, parity/drift_detection.py, parity/compliance.py) and the tests/ suite. See bob_sessions/A_plan_v2enhanced.md for the full multi-phase implementation transcript.
bob-the-judge/
├── README.md
├── LICENSE
├── requirements.txt
├── launch.py # Python launcher (all 3 services)
├── BobTheJudge.exe # Windows one-click launcher
├── dashboard.py # Streamlit UI (IBM Carbon Design System)
├── mcp_server.py # FastMCP server (4 tools)
├── bob-mcp-config.json # Bob IDE MCP registration
├── DEMO_SCRIPT.md # 3-minute demo script (word-for-word)
├── PITCH_SCRIPTS.md # 30s / 2min / 5min pitch scripts + judge Q&A
├── assets/gavel.svg
├── services/ # FastAPI legacy + modern banks (legacy uses COBOL via GnuCOBOL)
├── parity/ # parity engine, scoring, tenant_config, confidence, drift_detection, compliance
├── config/ # tenants.json (4 reference bank tier profiles)
├── audit/ # ReportLab PDF builder
├── bob/ # Bob client (watsonx.ai Granite + Bob REST + mock fallback)
├── tests/ # pytest suite (48 tests, 100% passing)
└── bob_sessions/ # IBM Bob exported sessions + consumption screenshots (5 sessions)
- Python 3.11+
- FastAPI 0.115 + Uvicorn
- Streamlit 1.39 + Plotly 5.24
- ReportLab 4.2
- pandas 2.2 + scipy + numpy (statistical methods)
ibm-watsonx-aiSDK (Granite inference)- mcp >= 1.0 (FastMCP)
- starlette < 1.0 (FastAPI compatibility pin)
- pytest + pytest-cov (test suite)
- IBM Plex Sans / Mono (UI typography via Google Fonts)
- GnuCOBOL (optional, real COBOL fee calculator in
services/cobol/FEE_CALC.cob)
BUILD FOR LABLABAI IBM BOB HACKATON 2026 SUBMISSION BY CYBERFALCON TEAM
MIT.