AI Analytics

Live · prod-us-east-1

Model performance, inference infrastructure & cost intelligence across every deployed AI system

Last synced 12 seconds ago

Generate AI Report

1.8pt

AI Accuracy (weighted avg.)

96.4%

0% Target 95% 100%

1.4pt above target · trending up

Classification

97.8%

Generative

94.9%

Multimodal

95.6%

7-day accuracy trend +1.8pt

Updated 2 minutes ago · 30-day rolling window

2.4pt

Model Performance Score

91.2/100

7-day trend

4 running

Training Jobs (30d)

182

129 done 31 running 22 failed

6.1%

GPU Cluster Usage

76.3%

Across 4 clusters · 32 GPUs

Inference Requests

8.92M/day

+14.2% vs last week

6,214 req/min avg.

AI Insights

Generated today, 09:14 AM · based on last 30 days of platform telemetry

Risk
Score

Confidence
Score

Platform-wide accuracy climbed to 96.4% this month, led by the support-tone-v3 fine-tune rollout. However, GPU cluster node-01 is running at 92% utilization with rising thermals, and Llama-3.1-70B shows a 2.3pt accuracy regression since the last dataset refresh. Token spend is trending to exceed the monthly budget by Aug 3 unless routing is optimized. Overall system risk remains low, and confidence in current model rankings is high given consistent evaluation coverage.

Reroute Vision traffic off node-01 before thermal throttling impacts latency SLA.

Shift summarization traffic to Claude-Haiku-4 to cut costs by ~$620/mo with minimal quality loss.

Consider rolling back Llama-3.1-70B to the pre-refresh checkpoint pending root-cause review.

Model Leaderboard

By usage share

1 OpenAI · GPT-4o-mini-ft-v352%
2 Anthropic · Claude-Sonnet-4.531%
3 Google · Gemini-1.5-Pro11%
4 Meta · Llama-3.1-70B6%

Top by accuracy

Claude-Sonnet

Top by speed

GPT-4o-mini

Top by cost

Gemini-1.5

Inference Volume

Requests per day, last 14 days

14.2% WoW

Token Usage Breakdown

By workload type, this month

Training Job Status

Jobs by outcome, last 6 weeks

Week	Done	Failed	Running
W1	12	2	3
W2	15	1	4
W3	18	3	2
W4	14	2	5
W5	19	1	4
W6	22	2	5

Avg. job duration 3h 12m

Model Latency (p50)

Milliseconds per request

GPT-4o-mini184ms

Claude-Sonnet312ms

Whisper-lg1,240ms

Llama-3.1-70B402ms

Embed-3-lg44ms

Fastest model

Embed-3-lg

Slowest model

Whisper-lg

Training & Fine-Tuning Jobs

31 running

Job Name	Base Model	Progress	Started	Status
support-tone-v4	GPT-4o-mini	78%	2h ago	Training
legal-doc-summarizer-v2	Claude-Haiku-4	Queued	—	Queued
churn-risk-classifier-v1	text-embedding-3-large	100%	1d ago	Complete
sentiment-tuned-v3	GPT-4o-mini	Failed	3h ago	Failed · OOM
onboarding-assistant-v1	Claude-Sonnet-4.5	34%	48m ago	Training
fraud-detector-v6	Gemini-1.5-Pro	100%	2d ago	Deployed

GPU Usage by Cluster

A100 · node-0192%

A100 · node-0278%

H100 · node-0361%

H100 · node-0444%

Vector DB & Knowledge Base

Vectors

42.7M

Namespaces

Index size

218 GB

Recall@10

98.2%

Last reindex: 14 min ago

Alerts & Anomalies

3 new

GPU node-01 thermal throttling risk

92% utilization · 2 min ago

Llama-3.1-70B accuracy regression

-2.3pt since dataset refresh · 1h ago

Token spend trending over budget

Projected overage by Aug 3 · 3h ago

Critical

Warning

Resolved (7d)

Inference Error Rate

Last 14 days

0.3pt

0.8%

down from 1.8% · 2 weeks ago

Timeout errors0.4%

Rate limit errors0.2%

Validation errors0.2%

Most affected Whisper-lg · 1.6%

Total errors (14d) 1,048 requests

Retry success rate 94.6%

Recent AI Conversations

1.2K today

Jordan D. · Support bot 2m ago

"Can you walk me through the refund process for..."

Resolved

Aisha M. · Onboarding assistant 9m ago

"How do I connect my Stripe account to the..."

In progress

Raj K. · Billing assistant 18m ago

"My invoice total doesn't match what was quoted..."

Escalated

RAG Retrieval Pipeline

1.1pt

Query funnel · last 24 hours · 214,880 queries

Queries received214,880

100% of total volume

Chunks retrieved (top-k)203,340

-5.4% dropped · empty index results

Passed relevance threshold184,910

-9.1% dropped · below similarity cutoff

Used in final answer176,220

82.0% end-to-end retrieval yield

Recall@10

98.2%

Avg. chunks/query

6.4

Index freshness

11m

Agentic Task Automation

312 runs today

Multi-step agent runs by outcome, last 24 hours

81%

Completed without human help253

Escalated to human review37

Failed · tool error22

Refund lookup & approval agent 96% success

Order status & tracking agent 74% success

Avg. steps/run

4.7

Avg. run time

38s

Cost saved (24h)

$1,240