Weflayr: the Margin Optimization Layer for AI builders

The Weflayr platform

Free up to 1B tokens ingested / month

Get full control of your unit economics

Weflayr captures every LLM call and joins cost to revenue, so you know which customers make you money and which ones cost you.

and then

Optimize your margins where it matters

See the margin gain from cheaper models and prompt-caching mechanisms, and act on it for the customers draining your margins.

Margin & savings overview Revenue Cost Margin Gross margin %

Total revenue

$831.12k

AI cost

$468.35k→$374.05k

−20.1%

Total margin

$362.77k→$457.07k

+26.0%

Gross margin

43.6%→55.0%

+11.4 pts

Role-play cost

$146.98k→$52.68k

−64.2%

P&L per customer2 customers at risk on margin

CustomerRevenue ↓CostMarginMargin %MoMStatus

Lumen Finance$301,691$98,653$203,03867.3%▲Healthy

Role-play simulation$37,033

Objection coach$16,313

Deal brief generation$19,684

Call scoring$13,524

Meeting recap$6,686

Knowledge Q&A$5,413

Northwind Labs$268,114$98,398$169,71663.3%▼Healthy

Role-play simulation$36,938

Objection coach$16,271

Deal brief generation$19,633

Call scoring$13,489

Meeting recap$6,669

Knowledge Q&A$5,399

Cobalt AI$122,683$46,251$76,43262.3%→Healthy

Role-play simulation$17,362

Objection coach$7,648

Deal brief generation$9,228

Call scoring$6,340

Meeting recap$3,135

Knowledge Q&A$2,538

Tessa Health$91,337$45,394$45,94350.3%▼Watch

Role-play simulation$17,040

Objection coach$7,506

Deal brief generation$9,057

Call scoring$6,223

Meeting recap$3,076

Knowledge Q&A$2,491

✦ Weflayr optimizationDrift Studios$58,900$46,531$12,36921.0%▲At risk

Role-play simulation$17,467

Objection coach$7,694

Deal brief generation$9,284

Call scoring$6,379

Meeting recap$3,154

Knowledge Q&A$2,553

✦ Weflayr optimizationBeacon Edu$41,200$42,436-$1,236-3.0%→At risk

Role-play simulation$15,930

Objection coach$7,017

Deal brief generation$8,467

Call scoring$5,817

Meeting recap$2,876

Knowledge Q&A$2,328

Illustrative data from the Weflayr product

How it works

Observe. Alert. Recommend. Optimize.

Observability is the way in. Optimization is the product.

01

Observe

One line of code captures every LLM call and joins cost to revenue. Live P&L per customer and feature.

02

Alert

Get told when a feature cost climbs or a customer turns loss-making. Before the bill lands.

03

Recommend

Cheaper models that hold quality or prompt caching optimization. Measured on your real traffic, with the margin gain measured before you ship.

04

Optimize

Enforce the recommendations.

Prompt caching optimisation Live

Weflayr's unique prompt caching optimisation engine indexes every prompt your project sent and finds the blocks that repeat, across calls, across features, across customers. It searches the optimal caching configuration your provider actually supports, restructures the prompt when the stable content sits in the wrong place, and validates each candidate on traffic it was not derived from.

Est. savings

$4.14k/mo

−15% of feature cost

Caching captured

61%

of achievable caching

A Prompt structure only anthropic/claude-opus-5 · Hoist the shared instructions above both volatile blocks Cross-feature $2.68k/mo

The features Support Chat and Auto-Summarize both open with a run_context header, and both re-inject a per-workspace activity block between the instructions and the worked examples. Two volatile blocks therefore sit inside the 5,560 tokens the two features send byte-identical, so the cache breaks at the first of them and nothing below is ever reused. Hoisting the instructions, the worked examples and the output format into one contiguous region, with both volatile blocks below them, lets the two features read a single cache entry instead of paying twice for the same tokens.

RISK: SAFE

The moved blocks reference nothing in the run header or the activity block, so both orderings read identically to the model. No wording changes.

B Config only anthropic/claude-opus-5 · Put the breakpoint at the end of the worked examples, and hold the cache an hour $1.46k/mo

The breakpoint sits after the first user turn today, so the cached region includes content that changes on every call and the entry is rewritten instead of read. Placing it at the end of the worked examples caches only the stable region. Support Chat's calls also arrive further apart than five minutes, so the hour TTL earns back its higher write price.

explicit · 5 min → explicit · 1 h

RISK: SAFE

Breakpoint placement and TTL are caching config. The prompt text and its order are untouched.

Finding A · one sampled call of the feature Support Chat, as it ran and with the fix

CachedNot cachedMoved by the fix

AS IT RAN · ~0 OF ~8,690 TOK CACHED

[system]

run_context: {"run_id": "8f3c1a94", "started_at": "2026-06-14T09:12:41Z", "locale": "en-GB"}

You are the Support Chat assistant for Northwind Labs. Answer only from the workspace's own documentation and ticket history; when the answer is not there, say so and offer to open a ticket. Never invent an API field, a plan limit or a price.

## Recent activity for this workspace last_export: 2026-06-13T22:04Z · open_tickets: 7 · last_seen_agent: rosa.k

## Worked examples Q: Can I raise the export row limit? A: It is 500k rows per export on enterprise [docs/exports#limits]. Split the range, or ask me to open a ticket. Q: Does deleting a workspace delete its exports? A: Exports live 30 days past deletion [docs/retention#exports], then go for good. … 6 more worked examples

## Output format Open with the answer, then the doc section in brackets. Ask at most one follow-up question, at the end.

WITH THE FIX · ~5,560 OF ~8,690 TOK CACHED

[system]

You are the Support Chat assistant for Northwind Labs. Answer only from the workspace's own documentation and ticket history; when the answer is not there, say so and offer to open a ticket. Never invent an API field, a plan limit or a price.

## Worked examples Q: Can I raise the export row limit? A: It is 500k rows per export on enterprise [docs/exports#limits]. Split the range, or ask me to open a ticket. Q: Does deleting a workspace delete its exports? A: Exports live 30 days past deletion [docs/retention#exports], then go for good. … 6 more worked examples

## Output format Open with the answer, then the doc section in brackets. Ask at most one follow-up question, at the end.

run_context: {"run_id": "8f3c1a94", "started_at": "2026-06-14T09:12:41Z", "locale": "en-GB"}

## Recent activity for this workspace last_export: 2026-06-13T22:04Z · open_tickets: 7 · last_seen_agent: rosa.k

Model benchmarking Live

Weflayr replays your real production runs on candidate models and judges each output against the one you actually shipped. You get the quality delta and the cost delta before you switch a single token.

Est. savings

$6.9k/mo

−74% of feature cost

Same or better

96.7%

4.1% better · 92.6% same · 3.3% worse

1,000 production runs replayed of the feature Support Chat · 94% of them ran on openai/gpt-5.5 at $0.0042 / run

Candidate modelQuality vs gpt-5.5Cost / runLatency

claude-haiku-4.5Anthropic

41 better · 926 same · 33 worse

$0.0009-79%

▼38%

Recommended

deepseek-v3.2DeepSeek

29 better · 917 same · 54 worse

$0.0006-86%

+12%

gpt-5-miniOpenAI

22 better · 911 same · 67 worse

$0.0004-90%

▼31%

ministral-3-14bMistral

12 better · 605 same · 383 worse

$0.0007-83%

▼19%

Quality drop

gemini-3.1-flash-liteGoogle

8 better · 583 same · 409 worse

$0.0003-93%

▼44%

Quality drop

Margin-aware semantic routing On the roadmap

Point your calls at Weflayr's gateway and the recommendations you accepted apply themselves on every request: the cheaper model where quality holds, the prompt order that hits the cache, the right path for each customer's margin.

Routes on live margin, not list price
All frontier models plus self-hosted open source in the EU
Uses the per-customer margin labels only you hold

Read the product docs

Integrate in minutes

Few lines to instrument. Every call captured.

Weflayr tracks all your LLM calls with one line of code. Then tag each call with the customer and the feature, and cost lands in your dashboard joined to revenue.

Read the technical docs

import weflayr

# one line: instrument your provider SDK, call sites unchanged
weflayr.auto_instrument(ai_sdks=weflayr.AiSdk.OPENAI)

# tag each call with who it is for and why
with weflayr.propagate_metadata(
    feature_name="support-chat", customer_id="c_123"
):
    client.chat.completions.create(
        model="gpt-5.5",
        messages=[{"role": "user", "content": prompt}],
    )