Architecture
The whole stack, end to end.
A Chromium extension + FastAPI backend that inspects every paste headed to ChatGPT, Claude, Gemini, Copilot or Perplexity. Veea Lobster Trap classifies, Gemini 2.5 rewrites, every event becomes a regulator-readable audit entry.
Every paste flows: employee → Conduit inspect → modal → audit log. Pick a scenario to step through.
Lobster Trap matched rule block_aws_access_key. Severity critical.
AKIA[0-9A-Z]{16} matched at LT layerRule redact_email_addresses · categories customer_pii · severity high
Alice Smith,alice@acme.com,$240K Bob Jones,bob@acme.com,$180K Carol Liu,carol@acme.com,$95K David Park,david@acme.com,$310K
Person A,user-a@example.com,$240K Person B,user-b@example.com,$180K Person C,user-c@example.com,$95K Person D,user-d@example.com,$310K
UI type crm_dashboard · severity high · categories customer_pii, financial_data
name, email, ARR Alice Smith, alice@…, $240K Bob Jones, bob@…, $180K Carol Liu, carol@…, $95K
I have a CRM dashboard showing 3 customer rows with columns name, email, ARR. Can you write a Python function that flags accounts with ARR > $200K?
"Engineering users attempted 3 critical credential exfiltration events in the last 24h, all blocked at the Lobster Trap layer. Finance users redacted 8 events containing customer PII, the largest cluster pasted into ChatGPT. Recommend reviewing the 2 policy overrides triggered by user u_8a7c."
Every part of the governance loop touches Gemini — text, vision, reasoning, search grounding, agentic function calling. All routed through Veea Lobster Trap, enforced by a CI test.
gemini-2.5-flash · JSON10 categories, severity, regulatory concern, specific findings — on every paste.
gemini-2.5-proRealistic placeholders, intent preserved. Employee still gets a useful answer.
flash + imageScreenshot pastes — the 2026 exfil vector nobody else catches. OCR + UI classification.
gemini-2.5-proWhen a screenshot is blocked, Gemini drafts a paste-safe text rewrite.
pro · thinkingAmbiguous cases reasoned about. Trace surfaced in the audit detail.
+ google_searchLive rotation steps, recent breaches, threat-actor TTPs — every claim cited.
function callingGemini investigates the audit log via Conduit's tools, then writes the brief.
embedding-001k-NN clustering: "this pattern appeared three times last week."
Conduit reframes Lobster Trap's role: not "filter prompts to my agent" but "filter prompts to the world's agents." Same engine, 100× the market. 24 rules, ingress + egress, network policy, regulator-readable audit log.
Eight distinct Gemini 2.5 surfaces, all via the OpenAI-compatible endpoint, all through Lobster Trap. Multimodal vision, thinking mode, search grounding, function calling, embeddings — every Gemini capability in scope, none decorative.
docker-compose up -d
brings Lobster Trap, backend, dashboard, and MCP server up from a clean clone. Load the extension unpacked, open the dashboard, paste anything into ChatGPT.