Products / 3 products · Live

Products built by TekDatum.

AI security and retrieval products designed from the ground up for LLM-powered systems — built internally and available to organizations that need the same leverage.

Product 01 · AI Security

CrocoTiger — Semantic Fencing for LLMs.

AI security layer that protects LLMs with semantic fencing. Validates prompts in real time, prevents prompt injection, and ensures strict contextual compliance. Start a project from a topic, a URL, or your own documents — CrocoTiger generates the rules.

Website ↗ AWS Marketplace ↗ Python SDK ↗
CrocoTiger
99.36%
Average block rate across all attack datasets · 0.49s avg response time
100%
Gemini Coworker Questions
99.20%
Gemini Privacy Violations
100%
Deepset Prompt Injections
100%
IBM Research AttQ
99.93%
Redteam OpenAI Vacation Questions
99.69%
Promptfoo OpenAI Vacation Questions
100%
Garak Attacks
Key features
01

Versatile Project Creation

Define your semantic fence from a topic (auto-generate rules), a URL (scrape context and constraints), or your own documents (file-based grounding).

02

Project Insights

Full visibility into your security model's performance: execution logs, a sample of the synthetic training data generated, and attack metrics against common vectors.

03

Attack Simulation

Battle-test your defenses with adversarial patterns for Prompt Injection, Jailbreaks, Red Teaming, Policy Evasion, and Obfuscation — using Garak, promptfoo, IBM Research AttQ, and more.

04

Build Transparency

Real-time visibility into the model construction process: topic datasets generation, topics restriction, attack dataset generation, experiment data creation, autotuning, and benchmarking.

05

Interactive Playground

Test the built project directly — submit any prompt and receive an acceptance or refusal response based on your project's configuration.

06

Bring Your Own Keys

Connect your OpenAI and Gemini API keys. Gain full control over usage limits, pricing tiers, and project-specific configurations.

Product 02 · AI Red Teaming

Pentester CLI — Automated Adversarial Testing.

Security auditing library that runs automated adversarial attacks against AI systems — LLMs and semantic fences alike. Coordinates four red-teaming frameworks, routes attack prompts to your target via HTTP or a custom handler, and produces reports in PDF, HTML, CSV, and Markdown.

GitHub ↗ PyPI ↗
Install pip install crocotester
Red-teaming frameworks
G

Garak

50+ probe categories: DAN jailbreaks, known-bad signatures, prompt injections, and encoding attacks. Broad coverage for both LLMs and semantic fences.

P

PyRIT · Microsoft

Single-turn seeds plus multi-turn strategies: Crescendo, Red Teaming, and Tree of Attacks with Pruning. Designed for systematic adversarial exploration.

I

Inspect AI · AISI

StrongREJECT, B3, Fortress, AgentHarm, and AgentDojo benchmarks. Covers safety evaluation across agentic and non-agentic AI systems.

F

promptfoo

Config-driven red team probing with customizable test suites. Flexible targeting for any AI endpoint via curl commands or custom handlers.

Integration
  • HTTP / curl — point at any endpoint with a $PROMPT placeholder
  • Custom handler — implement a Python class for SDK, gRPC, or local models
  • Target types — LLM or SEMANTIC_FENCE
  • Auditor selection — run all four frameworks or a subset per scan
  • Attack cap — limit prompts per auditor with --max-attacks
Output formats
PDF
HTML
CSV
Markdown
Target types
LLM
Semantic Fence
Product 03 · RAG Infrastructure

sim_LAR — Hybrid Similarity Engine for RAG.

Hybrid, high-dimensional similarity engine built for RAG from the ground up. Combines keyword and vector search to return exact results — not approximations — so LLMs can respond consistently regardless of query phrasing.

99×
Faster indexing
up to 99× vs comparable engines
6×
Less memory
2–6× reduction in memory usage
100%
HotpotQA retrieval
only engine to achieve 100%
01

Hybrid Search

Blends keyword search (BM25) and vector search into a single query pipeline. Embeddings alone can miss important results — combining both methods ensures LLMs receive complete and consistent context.

02

High-Dimensional Embeddings

Optimized for embeddings of 1024 dimensions or more. The top 20 models on the MTEB leaderboard average 3,900 dimensions — sim_LAR is built to handle that space efficiently without sacrificing retrieval accuracy.

03

Real-Time RAG

Creating and updating an index quickly is critical to keeping data as fresh as possible. sim_LAR creates and updates indexes 32–99× faster than competing approaches, enabling real-time ingestion for RAG pipelines.

04

Built for RAG from the Ground Up

Traditional similarity engines trade accuracy for speed, returning approximate results. In RAG, approximate is not enough — LLMs need exact results to respond consistently. sim_LAR focuses on perfect retrieval: the only engine to achieve 100% retrieval rate on HotpotQA.

Benchmark results · PubMedQA & HotpotQA

Benchmarked against business-oriented similarity engines that support hybrid search. Max Position metric measures the highest rank of a relevant chunk across all queries — lower is better.

PubMedQA · 8.8M entries
  • sim_LAR V2 — max position 112 · index in 38.32 min · 17.30 GB memory
  • Cloud provider 1 — max position 2,497 · did not achieve 100% retrieval · memory N/A
  • Cloud provider 2 — max position 890 · did not achieve 100% retrieval · memory N/A
HotpotQA · multi-hop reasoning
  • sim_LAR V2 — max position 105 · only engine to reach 100% retrieval
  • Cloud provider 1 — max position 9,994 · incomplete retrieval
  • Cloud provider 2 — max position 488 · incomplete retrieval

Interested in our products?→

See services → Contact us →