Skip to content
000 AI TRAINER
VALUEARCTECHNOLOGIES Start a project
PRODUCT 02 · MACHINE LEARNING

AI Trainer Scored answers with citations, evaluated before a customer sees them.

IN PRODUCTIONYOUR CLOUD OR OURSOPEN-WEIGHT OPTION
trainer.valuearc.tech/eval EVAL RUN
Corpora
Playground
Evaluation
Guardrails
Deployments
Q What is the cancellation window for the Maldives package after the deposit is paid?
A
Free cancellation up to 21 days before departure. Inside 21 days the deposit is retained per clause 4.2.
supplier-terms.pdf p.4 policy-2026.md
Answer accepted · added to regression set
RUN #412 · 1,240 CASES
Faithfulness94%
Answer recall88%
Citation match91%
Refusal rate6%
12 runs · score trending up since guardrail v3
AI TRAINER — EVALUATION RUN #412
0
Modules in the platform
0
Scores computed on every case
0
Citations required before an answer ships
0
Customer data sent for provider training
WHY IT EXISTS

Measurement is what separates a demo from a deployment.

A prototype that answers well proves nothing about the next change. AI Trainer scores every prompt, corpus and model revision against a fixed set of real questions, so an improvement can be demonstrated run after run.

BEFORE
01 Prompt changes judged by whoever read the output that day
02 No way to tell which document produced an answer
03 Regressions found by a customer before any test caught them
04 Model swaps treated as a leap of faith
AFTER
01 Every change scored against a fixed regression set
02 Citations rendered inline, traceable to page and file
03 Failing cases block the deploy like any other test
04 Model and cost compared on the same evidence
⇤⇥
MODULES

What the platform includes.

01 Corpora

Ingest PDFs, documents, tickets and site content. Chunking stays inspectable and re-runs when a source changes.

Chunk previewVersionedRe-index
02 Playground

Ask, compare two prompts side by side, and promote the stronger one into the evaluated set.

A/B promptsDiff viewPromote
03 Evaluation

Faithfulness, recall and citation match computed per case across the regression set, failures listed first.

Scored runsFailure triageHistory
04 Guardrails

Refusal rules, PII filters and escalation to a human below the confidence threshold you set.

ThresholdsPII filterEscalation
05 Deployments

Ship a version, watch its scores in production, and roll back to the last passing run in one action.

VersioningRollbackAlerts
06 Cost & latency

Per-model cost and p95 latency reported alongside quality, so the trade-off is visible at the point of choosing.

p95 latencyToken costModel mix
MODULE 01OF 06
01
Corpora
STEP 01

Ground it in your documents

Point it at the corpus. Chunking stays visible, so an answer that cites page four can be checked against page four.

STEP 02

Score every change

Every prompt, corpus or model change is scored per case against the regression set. Failures come back as a worklist.

STEP 03

Ship behind a guardrail

Below the confidence threshold the system refuses or escalates. Production scores appear on the same dashboard as the test runs.

trainer.valuearc.tech/eval EVAL RUN
Corpora
Playground
Evaluation
Guardrails
Deployments
Q What is the cancellation window for the Maldives package after the deposit is paid?
A
Free cancellation up to 21 days before departure. Inside 21 days the deposit is retained per clause 4.2.
supplier-terms.pdf p.4 policy-2026.md
Answer accepted · added to regression set
RUN #412 · 1,240 CASES
Faithfulness94%
Answer recall88%
Citation match91%
Refusal rate6%
12 runs · score trending up since guardrail v3
1Answer with inline citation
2Scored run · 1,240 cases
3Quality trend across versions
INTEGRATIONS & STACK

Model-agnostic by design.

Nothing in the pipeline is tied to one vendor. Switching a model is a configuration change followed by an evaluation run.

MODELS
Anthropic ClaudeOpenAIOpen-weight via vLLM
RETRIEVAL
pgvectorQdrantHybrid BM25
SOURCES
S3 / DriveZendeskConfluenceWeb crawl
STACK
PythonFastAPIPostgresAWS ap-south-1
QUESTIONS

What teams ask first.

No. Retrieval happens against your own store, and we run with training opted out at the provider. Self-hosted open-weight deployments are available where the policy requires it.

From your real questions. We take support history or a workshop transcript, write expected answers with your team, and that becomes the baseline everything is scored against.

It is logged as a failing case with the retrieved context attached and added to the regression set. The fix is verified on the next run.

Yes. The whole pipeline deploys into your AWS or Azure account, including the vector store and the evaluation service.

Start with a scored baseline.

Send a sample corpus and twenty real questions. We return a scored baseline and an estimate of what moving it costs.

Start a project→
REPLY WITHIN 48 HOURS
INDIA · MON–FRI