ZeroGrad
Capabilities · Recommendations

Product Recommendations

Healthy
In-context predictionEndpoint product-recommendationsProduction v12Challenger v13

Real-time production health

Known immediately — no outcomes needed.

Last 24 hours
Requests
782K
Availability
99.995%
P50 latency
24ms
P95 latency
58ms
Error rate
0.02%

Predictions / hour

20K33K46K00:0004:0008:0012:0016:0020:00

Latency

0ms54ms108msSLO P95 < 100ms00:0004:0008:0012:0016:0020:00
P50P95

Inputs, context & predictions

Schema
Healthy
Item catalog drift
Low
2.1% new items this week
Missing values
Normal
Cold-start requests
11.8%

Input drift vs reference (PSI)

0.1 · 0.25
item_category
0.05
price_bucket
0.04
session_length
0.03
device
0.02
user_tenure
0.01

Prediction distribution

today vs reference · no shift
0Top-1 recommendation score1
Live Reference examples

Routing

In-context prediction
v12 · control80%
v13 · experiment20%
Backend
Managed by Zero Grad
Context
pinned to v12
Fallback on timeout
Managed C · 0.02% of requests

Matured outcome quality

Performance metrics are computed only for predictions with observed outcomes. Clicks and purchases attributed within 7 days take up to ~7 days to arrive, so recent predictions aren't scored yet.

Evaluation window
Sep 17 – Sep 23
predictions made in this period
Outcome coverage
94%
have an observed outcome
Median outcome delay
1.6 days
clicks and purchases attributed within 7 days
Last outcome received
Oct 1, 09:12
via outcomes API

Business impact

Sep 17 – Sep 23 · 94% coverage
Revenue influenced
$412K
Orders with a rec click
9,860
Incremental revenue (est.)
$61K
vs. popularity holdout

Incremental revenue estimated against a 5% popularity-baseline holdout.

Model quality

Sep 17 – Sep 23
Online CTR
8.7%
Conversion
3.2%
NDCG@10 (replay)
.412
Catalog coverage
71%

NDCG@10 — matured outcomes

By prediction day. Recent cohorts are provisional until their outcomes arrive.

.409.412.416Sep 17Sep 19Sep 21Sep 23Sep 25Sep 27Sep 29
NDCG@10 · maturedProvisional (< 60% coverage)
Outcome coverage by cohorttoday →

Outcomes

Ground truth flows back by prediction_id. It improves evaluation and routing — never model weights.

Send outcomes

Every prediction returns a prediction_id. Send the real outcome whenever you learn it.

# Later, when the real outcome is known
zg.capability("product-recommendations").send_outcomes([
    {"prediction_id": "pred_01J9X4R3TC", "outcome": {"purchased": "sku_44120"},
     "observed_at": "2026-09-30T23:41:00Z"},
])

Outcomes never update model weights — Zero Grad does not train a customer-specific model. They are used to measure real production quality, compare foundation models and context-selection strategies, improve routing, and decide whether a challenger beats the current production configuration.

What Zero Grad learned from your outcomes

Simulated

No training doesn't mean no learning.

  1. Context length: the last 50 events match 200 events on matured purchases, at 40% lower latency.
  2. Cold-start users routed to the co-occurrence baseline: same CTR, 7× cheaper.
  3. Session events in context lift CTR on matured sessions (v13 experiment, day 4 of 14).

What adapts: which foundation model serves which slice, how in-context examples are selected, and calibration. What doesn't: model weights.