LogicGaze · SignalBuild 2026.11
  1. [01] mount trace store …
  2. [02] load span index · 155 traces …
  3. [03] warm evaluators · hallucination, relevance …
  4. [04] arm guardrails · 8 PII detectors …
  5. [05] connect providers · openai anthropic azure gemini …
  6. [06] render signal …
000
LogicGaze

LogicGaze · LLM observability and evals

Ship AI you can stand behind.

LogicGaze traces, scores, guards and costs every LLM call in production, so you know what your AI did, how well it did it, and what it cost.

Free: 10,000 traces a monthNo card neededSet up in minutes

[→] What LogicGaze is for

Works with OpenAI, Anthropic, Azure OpenAI and Google Gemini, as well as Mistral and Cohere.

[!!] The problem

Production AI
fails quietly.

A wrong answer, a leaked number, a prompt that talks the model out of its rules, a bill nobody expected. Without a record of every call, nobody can say which one did it.

46%of developers say they distrust the accuracy of AI tools.Stack Overflow Developer Survey 2025

  1. [01]Evaluate

    Wrong answers

    A confident hallucination reaches a customer.

  2. [02]Protect

    Leaked data

    A prompt carries an SSN or a card number to a model.

  3. [03]Protect

    Prompt injection

    A user talks the model out of its instructions.

  4. [04]Operate

    Runaway cost

    One deployment quietly eats the month’s budget.

  5. [05]Govern

    No record

    Nobody can say which call did it, or prove what happened.

Observe · LLM tracing

Know exactly what your agent did.

Every request becomes a timeline of agent steps, tool calls and model responses, with latency, tokens and cost on each span. Wrap your client once and every call is traced.

ExampleA research agent took 50.5s. The trace shows 18.2s of it went to a single gpt-4o call: synthesize-answer.

  • TracingA span tree for every request: latency, tokens and cost.
  • Agent graphMulti-step agent runs drawn as an execution graph.
  • Model performanceLatency percentiles, TTFT and error rates per provider.
  • DatasetsTurn production traces into golden datasets.
Send your first trace

Evaluate · LLM evals

Know when an answer goes wrong.

LLM-as-judge evaluators score live answers for faithfulness, relevance, coherence, hallucination and harmfulness, so quality becomes a number you can track, not a guess.

ExampleAn answer scores below your 0.50 threshold. It is flagged and sent for review, and its trace can become a regression test.

  • LLM-as-judgeFive dimensions, scored on live traffic.
  • Tests from real trafficAny production trace can become a test case.
  • Prompt managementVersioned prompts, compared side by side.
  • Human reviewAnnotation queues where a score is not enough.
Run your first eval

Protect · AI guardrails

Stop sensitive data at the gateway.

On the LogicGaze gateway, PII, prompt injection and harmful content are caught on the way in, blocked or redacted, and every event is logged.

ExampleA prompt contains a customer’s SSN. Your redaction rule replaces it before the model sees it, and the event is written to the audit log.

  • PII detection and redaction8 built-in detectors for SSNs, cards, passports, IBANs and more, with redaction rules you define.
  • Injection and jailbreaksBlocked on gateway requests before the model answers.
  • Evaluator guardrailsAny evaluator can flag failing responses just after they return.
  • Security scanning120+ SAST rules, 60+ secret patterns and dependency CVEs.
See guardrails in a demo

Govern & operate · LLM cost tracking

Prove what happened, and what it cost.

Every guardrail action, setting change and access decision lands in an audit log, next to token-level cost by model and provider, with alerts before a bad day becomes an incident.

ExampleAn evaluator flags a hallucination at 09:37. The audit log shows it beside the PII redaction at 09:42, and the cost view shows what that hour of traffic cost.

  • Audit logEvery enforcement action and admin change, recorded.
  • Retention and rolesRetention policies enforced daily; access by role.
  • Cost trackingPer-token cost across every provider, in USD.
  • AlertsError rates, latency spikes and cost overruns, by email, Slack or webhook.
See the cost view in a demo

[05] The product

The waterfall, the scores, the gateway and the cost curve above are drawn from the product. These are the screens themselves.

  • [01]TracingA span tree for every request: latency, tokens and cost.
  • [02]EvaluatorsFive dimensions, scored on live traffic.
  • [03]GuardrailsPII, prompt injection and harmful content, caught on the way in.
  • [04]Costs and audit logSpend by model and service, and a record of every change.

See it on your own traffic.

[06] Setup · Python SDK

  1. 01Install the SDK
  2. 02Call init once at startup
  3. 03Use your client exactly as before
app.py
pip install logicgaze
 
from logicgaze import init
import openai
 
init(api_key="lgz_...", base_url="https://app.logicgaze.com", auto_instrument=True)
 
# use OpenAI exactly as before: every call is now traced
client = openai.OpenAI()

Every call is traced with tokens, latency and cost, checked against guardrails and queued for evaluation.

[07] Security · data handling

The gateway sees prompts, so we are specific about what happens to them. Talk to us about your own data requirements.

Your keys, your path
By default the SDK calls your provider directly with your own keys. LogicGaze is not in the request path; the gateway is opt-in.
Redaction on the gateway
Eight built-in detectors flag PII such as SSNs, card numbers, passports and IBANs. Redaction rules you define replace matches before the prompt reaches the model.
Secrets masked
60+ secret patterns, such as AWS, OpenAI, GitHub and Stripe keys, masked before they reach AI or the UI.
Audit log
Every enforcement action, setting change and access decision is recorded.
Retention you set
Retention policies are enforced daily. The Free plan keeps traces for 14 days.
Access by role
Roles and permissions across your organization.

[08] Made for

  • 01See Observe

    AI engineers

    Every agent step, tool call and model response, with latency, tokens and cost on each span.

  • 02See Govern & operate

    Platform teams

    Cost, latency and error rates across every provider, compared from one place.

  • 03See Protect

    Security and compliance

    PII caught on the gateway, redacted by the rules you set, and every enforcement action written to an audit log.

  • 04See Evaluate

    Product leads

    Quality as a number: LLM-as-judge scores on live answers.

  • 05See Govern & operate

    Engineering leaders

    Alerts for error rates, latency spikes and cost overruns, before a bad day becomes an incident.

  • 06See Evaluate

    Review teams

    Annotation queues for the answers where a score is not enough.

[09] Pricing

Free

Everything you need to trace, score and guard production traffic.

$0/ month

  • 10,000 tracesevery month
  • 14-day retentionfor every trace
  • Up to 3 userson your team
  • No credit cardto get started

Paid plans scale capacity, retention and seats as your traffic grows.

[10] LogicGaze · FAQ

Questions, answered.

How LogicGaze fits into your stack, what it measures and what it costs to start.

What is LogicGaze?

LogicGaze is an AI trust platform for production LLM apps: it lets you see what your AI does, measure how good it is, stop it doing harm, and prove it. It covers LLM observability and tracing (Observe), LLM-as-Judge evaluation (Evaluate), guardrails and red teaming (Protect), audit logs and retention policies (Govern), and cost tracking, alerts, and gateway routing (Operate).

Which LLM providers does LogicGaze support?

LogicGaze works with OpenAI, Anthropic, Azure OpenAI, and Google Gemini, as well as Mistral and Cohere. You can compare cost, latency, and error rates across providers from a single dashboard.

How do I add LogicGaze to my app?

Install the SDK with `pip install logicgaze`, then wrap your client, for example `client = wrap_openai(OpenAI())`. Every call is traced automatically with no other code changes. By default the SDK calls your provider directly with your own keys, so LogicGaze is never in the request path and adds no latency to your LLM calls; routing through the LogicGaze gateway is opt-in. SDKs are available for Python and TypeScript, and most teams are set up in about five minutes.

What does LLM-as-Judge evaluation measure?

LogicGaze scores production outputs for faithfulness, relevance, coherence, hallucination, and harmfulness. You can turn real traces into test cases, compare prompt versions side by side, and add human review where you need it.

Can LogicGaze detect PII and prompt injection?

Yes. For requests routed through the LogicGaze gateway, rule-based guardrails detect PII with 8 built-in detectors (including SSNs, credit card numbers, passport numbers, and IBANs) and redact it with rules you define, block prompt-injection and jailbreak attempts, and filter harmful content. You can also promote any evaluator to a guardrail; these currently flag failing responses just after they are returned. Every enforcement action is recorded in an audit log.

Is there a free plan?

Yes. The Free plan includes 10,000 traces per month, 14-day data retention, and up to 3 users, with no credit card required. Paid plans scale capacity as you grow; see the pricing page for details.