Open Source · MIT License · v0.2.0

An 8-layer engine that turns raw events
into named friction patterns.

Velum takes raw JSON event streams and returns structured friction patterns with confidence scores and AI-generated hypotheses. Schema-free. One POST endpoint. Self-host in minutes.

POST /api/v1/analyze → response

$ curl -X POST http://localhost:8080/api/v1/analyze \
    -d '{"events": [...]}'

{
  "patterns": [{
    "type": "retry_storm",
    "severity": "critical",
    "confidence": 0.96,
    "affected_users": 847,
    "hypothesis": "2.3s btn_pay latency — no loading state visible"
  }],
  "status": "ok"
}

View on GitHub → Join Waitlist

MIT License · Self-hostable · One POST endpoint · Agent-ready

This is what your data looks like right now.

On the left: raw noise. On the right: what Velum extracts from it.

incoming events

12,847 events / min

1707500000 {"event":"checkout_page_view","user_id":"usr-101","device":"mobile"}

1707500002 {"event":"scroll","depth":340,"user_id":"usr-203"}

1707500003 {"event":"click","target":"btn_pay","user_id":"usr-101"}

1707500005 {"event":"page_view","path":"/pricing","user_id":"usr-445"}

1707500008 {"event":"click","target":"btn_pay","user_id":"usr-101","error":"card_declined"}

1707500010 {"event":"focus","field":"card_number","user_id":"usr-312"}

1707500012 {"event":"click","target":"btn_pay","user_id":"usr-101","error":"card_declined"}

1707500015 {"event":"idle","ms":45000,"user_id":"usr-445"}

1707500018 {"event":"page_view","path":"/pricing","user_id":"usr-445"}

1707500020 {"event":"blur","field":"card_number","user_id":"usr-312"}

1707500021 {"event":"exit","user_id":"usr-312","type":"abandon"}

1707500023 {"event":"click","target":"btn_back","user_id":"usr-445"}

1707500025 {"event":"click","target":"btn_pay","user_id":"usr-101","error":"timeout"}

1707500028 {"event":"page_view","path":"/features","user_id":"usr-445"}

1707500030 {"event":"click","target":"btn_back","user_id":"usr-445"}

1707500033 {"event":"exit","user_id":"usr-101","type":"rage_quit"}

1707500035 {"event":"page_view","path":"/pricing","user_id":"usr-445"}

1707500038 {"event":"exit","user_id":"usr-445","type":"abandon"}

velum pipeline →

detected patterns

4 patterns

🔥 Retry Storm Critical

usr-101 clicked pay 4× in 25s
error_code: card_declined → timeout
confidence: 0.96 · flow: checkout

🔄 Confusion Loop Warning

usr-445 looped pricing↔features 3×
45s hesitation between navigations
confidence: 0.89 · flow: evaluation

😶 Silent Abandonment Warning

usr-312 focused card field, paused, left
no error · no retry · session: 12s total
confidence: 0.91 · flow: checkout

AI Hypothesis

"Payment retry storm likely caused by 2.3s button latency — users don't see processing state. Confusion loop on pricing suggests hidden fee anxiety. Recommend: add loading indicator, surface fees on pricing page."

The problem

You're collecting events.
You're not extracting patterns from them.

📡

The events are there

Raw JSON streaming through your pipeline — Segment, Kafka, direct instrumentation. Hundreds of thousands per day.

> checkout_btn_click: 4,231
> conversion_rate: 3.2%
> avg_session: 4m 12s

🔍

But streams don't self-analyze

Retry storms, confusion loops, silent drop-offs — they're encoded in the event sequence. Writing detection logic for each one doesn't scale.

> custom query #1: retry detection
> custom query #2: loop detection
> custom query #47: ...

🔥

So someone writes a one-off script

An engineer queries the logs. The pattern recurs next sprint. No one connects the dots. The same friction ships again.

> Time to diagnosis: 2-4 weeks
> Pattern recurrence: every release
> Root cause found: rarely

What Velum Detects

Hidden friction your funnel metrics will never surface.

Critical

🔥

Retry Storm

Users hammering the same button — latency, missing feedback, or broken state.

usr-101: click click click click

Critical

😶

Silent Abandonment

Users arrive, see, and leave. No clicks. No scrolls. Just gone.

usr-312: view → ............ → exit

Warning

🔄

Confusion Loop

Back and forth between pages — uncertain, looking for answers they can't find.

usr-445: A→B→A→B→A→ exit

Warning

⏸

Hesitation

Long pauses before critical actions — doubt, anxiety, hidden friction.

usr-628: view_price → 47s pause → exit

Signal

🚪

Early Dropoff

Users abandon earlier than baseline — something broke or changed.

baseline: step 4 · actual: step 1

Signal

⏭

Bypass Behavior

Users skip expected steps — smart shortcut or broken flow?

expected: 1→2→3→4 · actual: 1→4

What makes Velum different

They tell you what happened.
Velum tells you why.

Amplitude · Mixpanel

PostHog · Hotjar

Velum

Event counting & funnels

✓

Session replay

—

✓

—

Friction pattern detection

—

✓

Explains why users struggle

—

manual

✓

Cross-user analysis at scale

✓

—

✓

Zero config / schema-free

—

✓

AI root cause hypotheses

—

✓

Self-hostable & open source

—

✓

Agent-ready API

—

✓

Velum runs alongside your existing stack — it doesn't replace it.
Feed it the same events you already track. Get back structured patterns and root cause hypotheses, not just counts.

How It Works

Eight layers. One API call. Real-time results.

Send raw JSON events. Get back structured friction patterns with AI-powered hypotheses.

Context Enrichment

Classifies unknown properties (error_code, plan_name) as targets or conditions.

AI-powered

Vocab Discovery

Tokenizes event names. Learns domain language. No config needed.

AI-powered

Event Normalization

Converts snake_case, camelCase, dot.notation into canonical form.

deterministic

Session Reconstruction

Re-assembles scattered events into coherent user journeys.

deterministic

Behavior Analysis

Tags each action as hesitation, retry, exploration, or abandonment.

deterministic

Friction Pattern Detection

Aggregates signals across users into named friction patterns with confidence scores.

deterministic

Baseline Comparison

Compares against historical snapshots. Flags what's new, increasing, or anomalous.

deterministic

AI Hypothesis

Generates natural language summaries with actionable explanations for each anomaly.

AI-powered

API-First

One endpoint.
Every agent.

Your AI agent POSTs events. Velum returns structured friction patterns, confidence scores, and AI hypotheses as JSON. The agent decides what to do next.

Step 1

Your Agent

POST raw event batches over HTTP. No SDK, no schema — any JSON works.

resp = requests.post(
"/api/v1/analyze",
json={"events": batch}
)
# → 200 OK, patterns inside

events

→

Step 2

Velum Analyzes

8-layer deterministic pipeline. Pattern detection, confidence scoring, AI hypothesis generation — all in one call.

8 layers · < 200ms · structured JSON

patterns

→

Step 3

Agent Acts

Parse the response. Route by severity. Trigger your downstream systems automatically.

p = resp["patterns"][0]
# severity: critical, users: 847
if p["severity"] == "critical":
jira.create(p)

🎫

File Tickets

jira · linear · github

jira.create_issue({
"type": "retry_storm",
"sev": "critical"
})

🚨

Trigger Alerts

slack · pagerduty · opsgenie

slack.post({
"text": "847 users affected",
"channel": "#oncall"
})

🎛

Toggle Features

launchdarkly · split · flags

ld.toggle({
"flag": "checkout_v2",
"enabled": true
})

📊

Feed Dashboards

grafana · datadog · prometheus

metrics.push({
"pattern": "retry_storm",
"confidence": 0.96
})

Deploy Your Way

Self-host it. Or let us run it.

Full control with self-hosting, or zero-ops with Velum Cloud. Same pipeline, same results.

Self-Hosted

Run on your infrastructure

Deploy Velum on your own servers. Your data never leaves your network. Full source access.

Go binary — single process, minimal dependencies
PostgreSQL for storage — use your existing DB
Bring your own LLM key for AI layers
Full data sovereignty — nothing phones home
Docker, Kubernetes, or bare metal

$ git clone https://github.com/gokulnair2001/Velum.git
$ cp example.config.yaml config.yaml
$ go run cmd/velum/main.go
✓ velum listening on :8080

Velum Cloud

Zero-ops, fully managed Coming Soon

In Development

We handle infrastructure, scaling, and upgrades. You send events, we return quantified friction patterns and prioritized healing recommendations.

No servers to manage — just an API endpoint
Auto-scaling to handle traffic spikes
Managed AI models — no API keys needed
Historical pattern storage and trend analysis
Deeper analysis with richer, more detailed results

Join Cloud Waitlist →

An 8-layer engine that turns raw events into named friction patterns.

This is what your data looks like right now.

You're collecting events.You're not extracting patterns from them.

The events are there

But streams don't self-analyze

So someone writes a one-off script

Hidden friction your funnel metrics will never surface.

Retry Storm

Silent Abandonment

Confusion Loop

Hesitation

Early Dropoff

Bypass Behavior

They tell you what happened.Velum tells you why.

Eight layers. One API call. Real-time results.

One endpoint. Every agent.

Self-host it. Or let us run it.

Run on your infrastructure

Zero-ops, fully managed Coming Soon

Deploy it.Own your data.

An 8-layer engine that turns raw events
into named friction patterns.

You're collecting events.
You're not extracting patterns from them.

They tell you what happened.
Velum tells you why.

One endpoint.
Every agent.

Deploy it.
Own your data.