Public Beta
Latence

Real-time protection for AI agents

TRACE protects AI agents from private-data leaks, unsupported answers, prompt attacks, drift, and wasted context - without replacing your stack.

View docs
  • Plug into your existing agents and pipelines
RAG question
Refund policy

Which refund terms apply to enterprise contracts?

Coding agent
Approval gate

propose_patch(refunds.ts) without approval gate

TRACE
Groundedness
0.45
Review
Privacy
Account number
Redacted
InfiniMem
62% dead weight removed
Decision
ReviewRepeatRelease

Enterprise AI scales with agents. So do their risks.

TRACE is for the moment an agent stops being a demo and starts touching customers, code, documents, and decisions.

Private Data Leaks

A support agent drafts from live customer records. One loose field can expose account data unnoticed.

Wrong Answers

RAG and coding agents sound confident even when the evidence is thin, stale, or never actually checked.

Wasted Context

Long runs drag old context forward, burn tokens, and bury the facts that should guide the decision.

Compliance

Missing logs and interpretability become attackable in audits and reviews.

TRACE sits beside the agent and checks what matters while the work is happening.

HOW TRACE WORKS

Drop TRACE next to your agent. Keep the control loop yours.

TRACE lives next to your RAG pipeline, coding agent, or tool-using workflow. It checks inputs, context, outputs, memory, and decisions in real time. Whether on autopilot or strict human-in-the-loop is up to you.

TRACE SDK
trace.pyreal API
result = trace.verify(
    answer=agent_answer,
    context=retrieved_context,
)

if result.decision == "review":
    route_to_human(result.evidence)

TRACE output

Trace wrong answers

decisionreview
groundedness0.45
unsupported_spans2
USE CASES

TRACE fits where agents already work.

Use the same protection layer for coding agents, support workflows, and legal review: grounded answers, efficient memory, private data handling, and audit-ready traces.

For long-running coding work

Detect drift over hundreds of steps

Detect task drift over hundreds of steps before the agent wanders away from the original goal.

Verify every claim against the codebase

Verify and score codebase-grounded reasoning at every step, including files, diffs, tools, and test claims.

Reduce bloat before it compounds

Catch unsupported claims and hallucinated output while reducing context bloat up to 90%.

QUALIFIED PILOT PHASE

Pilot before deployment. We deliver proven quality only. Period.

Before you pick the deployment, we prove TRACE on your data and workflows.

Cloud

For testing and low-risk workloads

Fast managed deployment for trace tests, demos, and non-sensitive evaluation.

Dedicated

Managed with stronger isolation

Predictable capacity and stronger isolation for early production teams.

VPC

Inside your cloud boundary

TRACE runs inside your cloud with your network, storage, and controls.

On-prem / air-gapped

For regulated environments

Full local control when production data cannot leave your environment.

Same API. Same SDK. Same product behavior. Different deployment boundary. Cloud is for evaluation and low-risk workloads; sensitive production data belongs in Dedicated, VPC, or on-prem deployments.

PRICING

Pay for the deployment. Not every trace.

You pay for the deployment. Throughput depends on the selected runtime. No surprise per-trace billing.

Flat monthly pricing by deployment size. No per-request billing. Throughput depends on selected deployment capacity.

Swipe plans

Sandbox

Evaluation
Free

Test TRACE with a few hundred traces.

  • Shared evaluation runtime
  • API + SDK access
  • Basic logs
  • Short retention
  • Not for production data

Starter

Shared cloud
€49/month

Shared cloud for development and low-risk testing.

  • Shared TRACE cloud
  • 5k-10k traces/month
  • RAG + coding-agent checks
  • PII redaction
  • Groundedness
  • Compression
  • 7-day retention

Pro Dedicated

Production
From €1,499/month

Dedicated managed deployment for production teams.

  • Dedicated TRACE backend
  • Dedicated runtime capacity
  • No per-request billing
  • Hardware-based throughput
  • Longer retention
  • Priority support
  • Custom configuration

Pro Small may be available from €799/month for lower-throughput dedicated deployments.

Enterprise

Private
Custom

Private deployment for sensitive and regulated workloads.

  • Customer VPC
  • On-prem
  • Air-gapped
  • Local logs
  • Custom retention
  • Security review
  • Deployment support

No sensitive production data belongs in shared evaluation paths. Production use starts with a qualified pilot and the right deployment boundary.

About Latence

Built by one focused founder Deployed with you.

I'm Dennis, the founder and builder behind Latence. If you test TRACE, you work directly with the person designing, shipping, and deploying the system.

No handoff maze. No consulting theater. Just direct work on the agent risks your team actually has.

Not for you. With you.

Book a callLinkedIn profile
Dennis Dickmann - founder of Latence
Dennis DickmannFounder · Latence
OPEN WORK

Built in the open where it matters

Latence is built by shipping real infrastructure, not slideware. Explore the SDK, retrieval experiments, and deployment work behind TRACE.

SDK

trace-sdk

Lightweight Python interfaces for calling TRACE from existing agent and RAG systems.

View SDK
Retrieval Engine

colsearch

Experimental infrastructure for future agent memory and retrieval optimization.

View project
Deployment stack

vllm-factory

Performance-focused serving experiments for private and high-throughput deployments.

See stack

Protect your agents before they reach production risk

Send a few real RAG answers or coding-agent traces. TRACE will show where private data, unsupported claims, drift, or wasted context appear.

Talk to Dennis
FAQ

What teams usually ask

The short answers most enterprise teams want before they start a pilot.

  • Is this another eval framework?
    No. Evals happen after the fact, on a dataset. TRACE sits beside the live agent workflow and returns a decision your system can act on: continue, repair, review, or block.
  • Do I have to switch my LLM or retriever?
    No. TRACE is designed to work beside your existing agent stack. You can keep your model, retriever, tools, orchestration framework, and deployment path.
  • What does TRACE check?
    Private data, prompt attacks, unsupported answers, context waste, drift, and auditability. Some guardrails are in private beta and are qualified during the pilot.
  • What does groundedness mean here?
    For every answer, TRACE estimates whether the supplied context supports the claim being made. The result is not just a score; it is a decision intent your application can route into continue, repair, review, or block.
  • Can we self-host?
    Yes, through a qualified design-partner path. Start hosted for testing, then move to Dedicated, VPC, or on-prem when prompts, code, customer data, or compliance requirements demand a private boundary.
  • What about my users' data?
    TRACE is designed around short-retention cloud evaluation and private deployment for sensitive workloads. Redaction, logging, and retention are qualified before a pilot handles production data.
  • How do I get started today?
    Use Run a trace test and send the workload you want to de-risk. We review the use case, deployment boundary, and data sensitivity before sharing the cleanest pilot path.
RESEARCH DNA

Algorithms and models are rigorously tested against industry standard benchmarks

BENCHMARKS

Fast enough for runtime

Built to score groundedness inline without turning your stack into a latency tax.

< 80 ms
p50 latency
118 ms
p95 latency

Benchmarked on finance, legal, QA, and multilingual data. Honest numbers at n = 60–120 per stratum.

Grounding Accuracypaired ranking
German tables
1.00
Finance prose
0.98
Semantic swap detection
0.98
Legal prose
0.96
Entity / date / number
0.95
Structured output
0.93
EN table-adversarial
0.88
RAGTruthF1 @ median
QA
0.69
precision 0.93
Summarization
0.68
Known Limitationstransparency
Data-to-text
0.45
HaluEval Dialogue ‡
0.57

‡ HaluEval Dialogue injects real-world facts absent from context — both answers score ungrounded. Does not affect RAG or agent use cases.

Data-to-text requires structured-to-prose transformation beyond current grounding scope.

All numbers from production config with NLI, reranker, and atomic claims enabled. n = 60–120 per stratum.