Evidence
Everything on this page is verifiable: diagrams, code excerpts and figures whose method is published.
This page does not claim — it shows. Every element here is verifiable: architecture diagrams, code excerpts from the deciding layer, and measurement figures whose method is stated.
1. Architecture: the request path through the hybrid WAF
The core design point: layers are ordered by processing cost, not by strength. Deterministic layers (rules and signatures) settle the majority, and the LLM arbitration layer only handles a small minority of suspicious requests. Cost and latency stay low while decision accuracy rises.
2. Excerpt from the arbitration layer
A short illustrative excerpt showing the decision logic: deterministic layers produce signals, arbitration is invoked above a threshold, with a timeout and a fallback path if the model does not respond.
# Arbitration layer - simplified logic
THRESHOLD = 0.62 # suspicion threshold from deterministic layers
TIMEOUT_MS = 350 # a slow model must never break the user experience
def decide(request, signals):
score = weighted_score(signals) # 0..1 from deterministic layers
if score < 0.20:
return ALLOW # most traffic ends here
if score > 0.85:
return BLOCK # decided without the model
verdict = llm_arbitrate(request, signals, timeout_ms=TIMEOUT_MS)
if verdict is None: # timeout expired
return CHALLENGE # challenge, do not block the unknown
return ALLOW if verdict.allow else BLOCK
The decision I consider most important here is the last line of the timeout branch: when arbitration fails we choose challenge, not block. Blocking on ignorance costs real users; a challenge costs an attacker one extra step.
3. Measurements — with the method stated
| Metric | Value | How it was measured |
|---|---|---|
| Requests reaching the arbitration layer | under 5% | Internal measurement on a test set mixing benign and attack traffic |
| Latency added by arbitration | none on 95% of requests | Response-time comparison with the layer disabled and enabled |
| False positives vs a fixed threshold | noticeably lower in design testing | A set of legitimate requests hardened step by step until a fixed threshold trips |
| Decision time for a suspicious request | capped at 350 ms | A hard bound fixed in code, not a stochastic result |
Transparency: the figures published here are internal measurements on test sets described in the third column, not production results for a client. Any number I did not measure myself is not published.
4. Evidence for the live tools
The seven tools actually run and can be exercised from a browser — the strongest evidence possible is real use, not a screenshot. Each tool has a project page explaining its scope and its evidentiary limits; all of them work on public information and bypass no authentication.
- TikTrace — public TikTok intelligence.
- GhostTrace — public Snapchat intelligence.
- X-OSINT — account and public activity analysis on X.
- Zehef — e-mail address checks linked to public sources.
- Tweet Recovery — retrieving publicly archived content.
- X Archive Pro — structured archiving.
- X Commander — formal report automation.
5. Written evidence: two full technical case studies
Two studies publish the methodology and the decisions rather than headlines only: the hybrid WAF study and the DDoS mitigation study. Each explains the problem, the options available, why each choice was made, and what did not work.
6. Runnable evidence: public repositories
Anything published under the MIT licence can be read, run and forked. Public account activity and the four repositories are on GitHub — with the publication standards shipped inside the repositories themselves.
7. What is not available yet — plainly
- Tool screenshots: the tools run live, and live use is more honest than a static image.
- Client production figures: the contracted projects ran under agreements protecting client data, so no attributable number is published.
- Credentials not yet earned: I do not publish a certificate before it is issued; plans and dates are public in the roadmap.