TECH.US AI RESEARCH

Tech.us AESM: An AI Model That Finds the Evidence Before It Answers

Meet the Agentic Evidence Search Model (AESM), a proprietary AI model developed by Tech.us to search enterprise knowledge, verify supporting evidence, and determine when there is enough information to answer reliably.

Built for the Tech.us Company Brain, AESM brings evidence-first AI into the organization's own infrastructure.

  • 83.6% Balanced Accuracy
  • 26.3-Point Lead Over the Best Hosted Model Tested
  • 98/98 Prompt-Injection Attempts Resisted in Testing

MODEL OVERVIEW

Meet AESM: The Evidence Engine for Enterprise AI

A compact, purpose-trained model that checks the evidence before an answer is generated.

AESM sits between enterprise search systems and the final answer. It identifies missing information, evaluates retrieved records, and decides whether to continue searching, proceed with an answer, or report insufficient evidence.

Retrieval Enterprise Search
Evidence Verification AESM

AESM decides to

Continue searching
Proceed with an answer
Report insufficient evidence
Output Verified Evidence & Citations
  • ~2 Billion Parameters Compact, purpose-trained model
  • Evidence Traceability Searches, citations, and decisions recorded
  • No Per-Token API Fees Runs on your own hardware.

BENCHMARK PERFORMANCE

83.6% Balanced Accuracy. Ahead of Three Leading Hosted AI Models Tested

Evaluated on 500 unseen questions using the same retrieval environment.

+26.3 pts

26.3 percentage points ahead of the leading hosted AI model tested.

AESM correctly resolved more than three times as many answerable questions as any of the three leading hosted AI models and correctly declined 95.9% of questions without sufficient evidence.

Accuracy comparison across Tech.us AESM and three leading AI models
Metric AESM GPT-6 Luna Claude Haiku 5.5 Gemini 3.1 Flash-Lite
Balanced Accuracy
83.6%
57.3%
55.5%
52.0%
Answerable questions resolved correctly (of 258) 184 71.3% 60 23.3% 54 20.9% 56 21.7%
Unanswerable questions correctly declined (of 242) 232 95.9% 221 91.3% 218 90.1% 199 82.2%
Total correct (of 500) 416 281 272 255

Source: Tech.us internal testing, 2026. Results are specific to this evidence-search benchmark.

THE ENTERPRISE CHALLENGE

Enterprise Search Finds Documents. AESM Checks Whether They Answer the Question.

Enterprise knowledge is spread across documents, tickets, wikis, and internal systems. Records can be incomplete, outdated, contradictory, or misleading.

General-purpose AI models may answer before gathering sufficient evidence or continue searching after the answer is available.

AESM addresses this gap by evaluating what the evidence supports before the answering process continues.

HOW AESM WORKS

From Enterprise Questions to Verified Evidence

Five connected steps designed to make AI answers traceable and evidence-supported.

  1. Step 1: Understand

    Breaks a question into required facts and constraints.

  2. Step 2: Search

    Plans searches across relevant enterprise knowledge sources.

  3. Step 3: Verify

    Checks supporting records for completeness, relevance, and conflicts.

  4. Step 4: Decide

    Continues searching when needed or flags insufficient evidence.

  5. Step 5: Record

    Preserves the search history, cited records, and decisions for review.

AESM works alongside existing enterprise search and retrieval systems. It does not generate the final answer itself.

SEARCH EFFICIENCY

More Accurate Evidence Decisions. Fewer Searches.

AESM reduced search activity and token processing in the same 500-question evaluation.

1,636

Total Searches

2.99

Average Model Calls per Question

2.30M

Input Tokens Processed

Compared with the Hosted AI models tested:

AESM search reduction
  • 46% fewer searches than GPT-6 Luna
  • 44% fewer searches than Claude Haiku 5.5
  • 29% fewer searches than Gemini 3.1 Flash-Lite

AESM also processed 42% to 76% fewer input tokens, reducing retrieval workload and compute requirements.

SECURITY AND CONTROL

Enterprise Evidence. Under Enterprise Control.

Built for traceability, prompt-injection resistance, and deployment within customer infrastructure.

Resilience Against Prompt-Injection

AESM followed none of the injected instructions across 98 benchmark cases covering four attack patterns. This does not imply immunity to all attacks.

Evidence-Linked Citations

Every cited record must support the specific fact being evaluated. AESM also maintained valid structured output throughout the benchmark.

Private Infrastructure Deployment

AESM runs inside the customer's environment on a single consumer-grade GPU, keeping internal queries and evidence checking under organizational control without external per-token API fees.

Your Enterprise Knowledge Deserves Evidence-First AI

Discover how AESM can support more reliable answers across your existing knowledge systems.

Discuss Your AESM Use Case

BUSINESS VALUE

What AESM Brings to Enterprise AI

Evidence-first decisions that support accuracy, efficiency, and control.

  1. Fewer Unsupported Answers

    Recognizes when available records cannot reliably answer a question.

  2. Efficient Cross-System Search

    Targets missing information without unnecessary retrieval steps.

  3. Predictable Operating Costs

    Runs on consumer-grade hardware without per-token API fees.

  4. Traceable Decisions

    Maintains records of searches, citations, and evidence assessments.

  5. Greater Data Control

    Operates within the organization's existing infrastructure and security environment.

COMPANY BRAIN

The Evidence Engine Behind Tech.us Company Brain

AESM is part of the Tech.us Company Brain, an AI knowledge layer designed to turn existing company records into answers grounded in verifiable evidence.

It is designed to work with knowledge sources such as Jira, Confluence, SharePoint, CRM systems, and internal communications.

Tech.us designs, deploys, and tunes the Company Brain for each customer's environment through its AI-augmented delivery model, supported by teams in the United States and India.

Knowledge sources

  • Jira
  • Confluence
  • SharePoint
  • CRM systems
  • Internal communications
  • & More...
Tech.us Company Brain Powered by AESM

RESEARCH METHODOLOGY

How We Measured AESM's Performance

A controlled evaluation of evidence retrieval, verification, and answer readiness.

Answerable Questions
258
Unanswerable Questions
242
Prompt-Injection Test Cases
98

All four models used identical questions and the same retrieval environment. The simulated company environments were excluded from AESM's training.

Balanced accuracy measures the average accuracy across answerable and unanswerable questions. Correct answers required every necessary fact to be supported by valid citations.

The benchmark covered:

  • Missing and conflicting information
  • Look-alike records
  • Restricted records
  • Multi-part questions
  • Prompt-injection attempts

Results are based on Tech.us internal testing completed in 2026 and may differ in other environments.

EXECUTIVE PERSPECTIVE

Why Tech.us Built AESM

Purpose-trained AI for a specific enterprise challenge.

“Enterprises need AI that can show its work and say plainly when the evidence is missing. AESM shows that a compact model trained for one job, finding and checking evidence, can outperform widely used general-purpose models at that job while running inside the company's own infrastructure.”
Praveen Narra Praveen Narra Founder & CEO, Tech.us

FAQ

Frequently Asked Questions About AESM

AESM (Agentic Evidence Search Model) is a proprietary AI model that plans enterprise searches, checks retrieved evidence, and determines whether enough information exists to answer reliably.

Traditional RAG retrieves documents for answer generation. AESM adds an evidence-verification layer that checks whether each required fact is supported before the answer is generated.

AESM works alongside existing enterprise search and retrieval systems, directing searches and evaluating their results.

AESM verifies whether sufficient evidence is available, while final answer generation is handled separately.

 Yes. AESM runs on a single consumer-grade GPU inside the customer's environment, without requiring external per-token API fees. 

It evaluates retrieved records, searches further when appropriate, and reports insufficient evidence when the available information cannot support a reliable answer.

AESM is a proprietary Tech.us model and is not publicly released. Organizations can contact Tech.us to discuss applying AESM to their enterprise knowledge systems.

Mid-market and enterprise organizations building internal AI assistants or a Company Brain, particularly those requiring reliable evidence, traceability, and data control.

Put Your Enterprise Knowledge to Work With AESM

Bring evidence-first AI into your organization's existing knowledge infrastructure with Tech.us Company Brain.

Talk to Tech.us About AESM