Appearance
GhostScience Demo Script
Eight demos. Twenty minutes. PhD-level science with full PII protection and regulatory provenance.
This script is designed for live customer demos, conference presentations, and internal stakeholder walkthroughs. Each demo is self-contained with exact inputs, expected outputs, and talking points.
Pre-Demo Checklist
Run through this before every demo:
- [ ] GhostAI instance is running and reachable
- [ ] PhD Mode is enabled:
PHD_MODE_ENABLED=truein.env - [ ] Sandbox is enabled:
SANDBOX_ENABLED=truein.env - [ ] PhD sandbox image is built:
docker images ghostai-phd-sandbox - [ ] Docker daemon is running:
docker info - [ ] API key is valid:
curl -s -H "Authorization: Bearer $GHOSTSCIENCE_API_KEY" https://your-host/api/phd/health - [ ] Health check returns
"status": "ok"and"sandbox_available": true - [ ] Browser is open to GhostAI chat interface
- [ ] PhD Science tab is visible in the empty state demo selector
- [ ] Screen resolution and font size are large enough for audience visibility
Quick health check
bash
curl -s -H "Authorization: Bearer $GHOSTSCIENCE_API_KEY" \
https://your-host/api/phd/health | python -m json.toolExpected:
json
{
"status": "ok",
"sandbox_available": true,
"cache_connected": true,
"cache_backend": "redis",
"phd_mode_enabled": true
}Demo 1: Symbolic Calculus
Duration: 2 minutes Story: A researcher needs to integrate a non-trivial expression and verify the result. GhostScience does it with step-by-step LaTeX and automatic verification.
Input
Type this into the GhostAI chat:
Integrate x² · sin(x) with respect to x and show me the full step-by-step derivation using integration by parts. Then verify the answer by differentiating.Expected Output
The AI invokes the symbolic_math tool. The result includes:
- LaTeX-rendered equation with the antiderivative: $-x^2\cos(x) + 2x\sin(x) + 2\cos(x) + C$
- Step-by-step derivation showing two rounds of integration by parts
- Verification showing that differentiating the result recovers $x^2\sin(x)$
- Provenance badge with computation ID and wall time
Talking Points
"GhostScience doesn't just compute the answer — it shows the full derivation, step by step, exactly like a PhD advisor would write on a whiteboard. Then it automatically verifies the result by differentiating back. Try getting MATLAB to explain its reasoning."
"Notice the provenance badge. Every computation is recorded with the tool name, parameters, generated code, and result hash. This is a regulatory-grade audit trail — it satisfies FDA 21 CFR Part 11 and SOX Section 404."
Demo 2: Multi-Syntax Input
Duration: 2 minutes Story: Different teams use different notation. GhostScience accepts all of them.
Input
Send these three messages in sequence:
LaTeX:
Evaluate $\int_0^1 x^3 \, dx$Natural language:
Integrate x cubed from 0 to 1MATLAB-style:
integral(@(x) x.^3, 0, 1)Expected Output
All three produce the same result: $\frac{1}{4} = 0.25$
Each response shows the detected syntax (latex, natural, matlab) and the normalized SymPy expression.
Talking Points
"A physicist writes LaTeX. A controls engineer writes MATLAB. A domain expert writes English. GhostScience accepts all of them. The DSL parser detects the syntax automatically and normalizes it to executable Python. Your team doesn't need to learn a new language."
"This is a real differentiator against MATLAB, which only accepts MATLAB syntax, and Wolfram, which only accepts Wolfram Language."
Demo 3: ODE System
Duration: 3 minutes Story: Solve a famous chaotic system and visualize the trajectory interactively.
Input
Solve the Lorenz attractor system (σ=10, ρ=28, β=8/3) from initial conditions [1, 1, 1] over t=0 to 50. Plot the 3D chaotic trajectory and the time series of each variable.Expected Output
The AI invokes numerical_solve with problem_type: "ode". The result includes:
- Interactive 3D Plotly chart of the butterfly attractor (zoom, rotate, hover)
- Time series plots of x(t), y(t), z(t) showing chaotic behavior
- Solver metadata: method used (RK45 or LSODA), step count, wall time
- Provenance record with environment hash
Talking Points
"This is a classic chaos theory problem — the Lorenz attractor. Notice the interactive 3D plot. You can rotate it, zoom in, hover over points. This is Plotly, not a static image. Try doing that in a MATLAB terminal."
"The solver automatically chose RK45 because it detected the system is non-stiff. If it were stiff, it would switch to BDF or LSODA. GhostScience classifies problem structure before selecting a solver — that's the AI reasoning layer."
"And every data point in this computation was PII-scanned on the way in and on the way out. If someone accidentally included patient IDs in their initial conditions, they'd be tokenized before reaching the sandbox."
Demo 4: Statistical Analysis
Duration: 2 minutes Story: Run a full regression analysis with diagnostics — the kind of work a biostatistician does daily.
Input
Generate a sample dataset of 200 points with a linear trend plus noise, then run a full linear regression analysis. Show R², p-values, confidence intervals, and generate QQ and residual diagnostic plots.Expected Output
The AI invokes statistical_analysis with analysis_type: "regression". Output includes:
- Regression summary table: R², adjusted R², F-statistic, coefficients with standard errors, t-values, p-values, 95% confidence intervals
- QQ plot (normality of residuals)
- Residual vs. fitted plot (heteroscedasticity check)
- Histogram of residuals
Talking Points
"This is what a data scientist spends their afternoon doing in R or Python — setting up imports, writing boilerplate, formatting tables. GhostScience does it in one sentence. And the diagnostic plots are interactive Plotly, not static PNGs."
"For healthcare and pharma, the provenance record for this regression satisfies FDA audit requirements. Every parameter, every data point, every method choice is recorded."
Demo 5: Optimization
Duration: 2 minutes Story: Minimize a notoriously difficult test function and visualize the solver's path.
Input
Minimize the Rosenbrock function f(x,y) = (1-x)² + 100(y-x²)² starting from (-1, 1) using L-BFGS-B. Visualize the optimization path on a contour plot and report the convergence history.Expected Output
The AI invokes optimize. Output includes:
- Optimal point: x=1.0, y=1.0, f(x,y)≈0.0
- Contour plot with the optimization path overlaid (showing the solver navigating the narrow curved valley)
- Convergence history (function value vs. iteration)
- Solver stats: iterations, function evaluations, gradient evaluations
Talking Points
"The Rosenbrock function is a standard torture test for optimizers — it has a narrow curved valley that's easy to find but hard to follow. Watch the path on the contour plot. L-BFGS-B navigated it efficiently."
"GhostScience auto-classifies optimization problems. If this were a convex problem, it would use CVXPy. Multi-objective? pymoo with Pareto fronts. The AI selects the solver — the user describes the problem."
Demo 6: Control Systems
Duration: 2 minutes Story: Design a controller and analyze stability — the daily work of a controls engineer.
Input
Design a PID controller for the transfer function G(s) = 1/(s² + 2s + 1). Show the Bode plot, step response, and gain/phase margins. Tune for minimal overshoot.Expected Output
The AI invokes control_systems. Output includes:
- Bode plot (magnitude and phase vs. frequency)
- Step response showing rise time, settling time, overshoot percentage
- Gain margin and phase margin values
- PID gains: Kp, Ki, Kd
Talking Points
"This is a second-order system — common in motor control, process control, robotics. GhostScience designed a PID controller, plotted the Bode diagram, and measured the stability margins. A controls engineer would normally need MATLAB's Control System Toolbox at $2,000/year for this."
"GhostScience uses python-control under the hood — the same transfer function algebra, but with AI-guided tuning and automatic stability analysis."
Demo 7: Provenance and Reproduce
Duration: 3 minutes Story: Show the regulatory audit trail and prove that results are reproducible.
Input
Step 1 — Run a computation (use a result from an earlier demo, or send):
Compute the eigenvalues of the matrix [[4, 1], [2, 3]] and verify they satisfy the characteristic polynomial.Step 2 — After the result appears, note the computation_id from the provenance badge. Then ask:
Show me the full provenance record for that computation, then export it to LaTeX.Step 3 — Ask:
Reproduce that computation and verify the result hash matches.Expected Output
- Step 1: Eigenvalues [5, 2] with verification
- Step 2: Full provenance record (tool, parameters, generated code, environment hash, result hash, wall time) + LaTeX document export
- Step 3: Reproduction result showing
matches: truewith both hashes identical
Talking Points
"This is where GhostScience is fundamentally different from every other scientific tool. We just proved, cryptographically, that re-running the exact same code in the exact same environment produces the exact same result. The hash matches."
"For FDA submissions, SOX audits, or HIPAA reviews, this is the gold standard. The provenance record shows who ran what, when, with what parameters, in what environment, and the result hash proves it hasn't been tampered with."
"The LaTeX export generated a publication-ready document with the computation method, parameters, and results. A researcher can drop this straight into a paper."
Demo 8: Method Comparison
Duration: 2 minutes Story: Compare multiple solvers on a hard problem — let the AI recommend the best one.
Input
Compare RK45, BDF, and LSODA numerical solvers on the Van der Pol oscillator (μ=1000, stiff regime). Benchmark accuracy, step count, and wall-clock time. Show solution plots side by side.Expected Output
The AI invokes compare_methods. Output includes:
- Side-by-side solution plots for all three solvers
- Comparison table: accuracy metric, step count, wall time, memory
- Recommendation: "BDF is recommended for this stiff system because..."
- Rationale: explanation of why RK45 struggles with stiffness parameter μ=1000
Talking Points
"This is one of the most powerful features. GhostScience solved the same problem three different ways and gave you a head-to-head comparison with a recommendation. It detected the stiffness and explained why the explicit solver (RK45) needs far more steps."
"No other tool does this. MATLAB makes you run each solver manually and compare results yourself. GhostScience automates the entire benchmarking process."
Recovery Steps
Things go wrong in live demos. Here's how to recover:
PhD Mode disabled
GET /api/phd/health returns "phd_mode_enabled": falseFix: Set PHD_MODE_ENABLED=true in .env and restart the backend.
Sandbox not available
GET /api/phd/health returns "sandbox_available": falseFix: Verify Docker is running (docker info) and the PhD image exists (docker images ghostai-phd-sandbox). Rebuild if needed:
bash
cd infrastructure/docker/sandbox
docker build -f Dockerfile.phd -t ghostai-phd-sandbox:latest .Computation times out
The default PhD timeout is 300 seconds. For complex demos (Lorenz attractor, Monte Carlo), increase it:
PHD_SANDBOX_TIMEOUT=600Plotly charts not rendering
Verify the frontend is receiving structured JSON with plotly keys. If the chart appears as raw JSON, refresh the page — the Plotly renderer may not have loaded.
Wrong tool invoked
If the AI picks a different tool than expected, be explicit:
Use the symbolic_math tool to integrate x² · sin(x) with respect to x.Audience asks about latency
Typical wall times: symbolic math 1-3s, numerical ODE 2-5s, optimization 3-8s, Monte Carlo 5-15s. Point to the wall_time_ms in the provenance badge. The sandbox overhead is ~500ms for container startup; persistent sessions eliminate this.
Demo Flow Cheat Sheet
| # | Demo | Duration | Key Moment |
|---|---|---|---|
| 1 | Symbolic Calculus | 2 min | Step-by-step LaTeX derivation + verification |
| 2 | Multi-Syntax Input | 2 min | Same answer from LaTeX, English, and MATLAB |
| 3 | ODE System | 3 min | Interactive 3D Lorenz attractor |
| 4 | Statistical Analysis | 2 min | Full regression with diagnostic plots |
| 5 | Optimization | 2 min | Rosenbrock contour with solver path |
| 6 | Control Systems | 2 min | PID design with Bode plot |
| 7 | Provenance & Reproduce | 3 min | Audit trail + bitwise verification |
| 8 | Method Comparison | 2 min | Three solvers head-to-head with recommendation |
| Total | ~18 min |
Frequently Asked Questions (From Demos)
"How does this compare to MATLAB?"
MATLAB is $10,000+/seat/year for full toolboxes. GhostScience is included in GhostAI. MATLAB has no PII protection, no audit trail, no reproducibility guarantees, no AI reasoning layer, and only accepts MATLAB syntax. GhostScience accepts LaTeX, MATLAB, Python, and natural language.
"What about Wolfram Mathematica?"
Wolfram has excellent symbolic math. GhostScience is strong and improving (SymPy-based). But Wolfram has no PII protection, no provenance, no federated clean rooms, and no AI method selection. For regulated industries, the compliance gap is the deciding factor.
"Is the PII protection real or just marketing?"
Every input passes through the PreDact PII pipeline before reaching the LLM or sandbox. Every output is scanned before delivery. The sandbox has no network access and runs under gVisor. Show the provenance record — it includes the PII scan timestamp.
"Can I reproduce these results in 6 months?"
Yes. The provenance record stores the environment hash (Docker image + library versions), the generated code, and the RNG seeds. The reproduce endpoint re-executes in the same environment and verifies bitwise. This is the only scientific computing platform that guarantees this.
"What happens with sensitive data like patient records?"
PII in computation inputs is tokenized before reaching the sandbox. PII in outputs is caught before reaching the user. The computation runs on tokenized values. For healthcare, the provenance record includes HIPAA compliance tags.
"Does it work offline / air-gapped?"
Yes. GhostScience deploys via Zarf into disconnected environments. All 50+ libraries are pre-installed in the sandbox image. No runtime network access required.
For hands-on setup from scratch, see the GhostScience Full Lab. For federated clean room demos, see the Clean Room Demo Script.