Skip to content
@PSCRedefine

PSCRedefine

PSC Redefine — calibrated engagement prediction for short-form video

Python PyTorch LightGBM FastAPI DuckDB Tests Hiring

What we're building

Short-form feeds are ranked by models nobody can audit, validated offline on logs shaped by the very policy being replaced. PSC Redefine builds engagement prediction you can reason about: calibrated probabilities instead of opaque scores, evaluation that separates real signal from exposure bias, and serving with monitoring built in — not bolted on.

Our work runs on real production logs at scale: 322M interactions, ROC-AUC 0.9147, every figure reproduced from a committed artifact.

Research: RankShift · live page

KuaiRand's uniformly-random exposure window gives a genuinely missing-at-random test set — a rare gold standard. Using it, we isolate what survives a policy change and what silently breaks:

  • Ranking transfers; calibration does not. Top-of-traffic lift holds across exposure policies (3.43× → 3.36×), but predicted probabilities run 2.07× too high on randomized traffic — a model that looks fine offline ships probabilities wrong by a factor of two.
  • Off-policy estimators graded against ground truth. IPS, SNIPS, DM and DR all miss the uniform policy's true value (3.0–3.9× off) — item-level propensities can't express the user–video matching that drives most of the exposure bias.
  • Findings replicate at 322M rows, evaluated in full, out of core (DuckDB), where the calibration break grows to 2.29×.

The serving stack: RankShift Serving · plain-language tour

Service What it does
SinglePrediction Train, select, and serve — real-time scoring API, ~1 ms per prediction
BatchPrediction Bulk scoring with per-row fault isolation
ModelInfo Model artifact introspection API + health dashboard
AnalyticsDashboard Request logging, drift alerts, live monitoring

FastAPI + Streamlit per service, Docker Compose, leakage blacklists enforced by tests, chronological splits, paired-bootstrap model selection, operating points set against a traffic budget. 292 tests, CI on Python 3.11–3.13, MIT — and each repo states its known gaps in docs/PRODUCTION_READINESS.md.

How we work

  • Calibration is a feature. A probability should mean what it says — we measure calibration error and publish it.
  • Evidence prunes everything. Features, models, even our own earlier conclusions: the same screening rule that kept 2 of 25 features on synthetic data keeps 24 of 37 on real logs, and we documented why.
  • Reproducible by default. Every reported figure regenerates from a committed JSON artifact and a fetch script.
  • Monitoring ships with the model. Error-rate, latency, and output-distribution alerts live beside serving, not after it.

🚀 We're hiring

Founder-led, early-stage, and looking for our first hire: a Founding Engineer to own major parts of the platform end to end.

→ Read the role & apply

Founder

Mengyun Wang — ML/AI infrastructure engineer; previously training-data infrastructure at Meta and GPU/training software at Oracle. GitHub · Site · mengyun_wang_ai@outlook.com

PSC Redefine · building feeds you can reason about

Popular repositories Loading

  1. SinglePrediction SinglePrediction Public

    RankShift Serving — train, select and serve the engagement model: leakage-tested features, paired-bootstrap selection, budget-based operating point

    Python

  2. BatchPrediction BatchPrediction Public

    RankShift Serving — batch engagement prediction with per-row fault isolation

    Python

  3. ModelInfo ModelInfo Public

    RankShift Serving — model introspection API and health-check dashboard

    Python

  4. AnalyticsDashboard AnalyticsDashboard Public

    RankShift Serving — request analytics middleware, aggregation API and monitoring dashboard

    Python

  5. .github .github Public

    Organization profile: what the Cognitive Shorts repositories are and the order to read them in

  6. rankshift rankshift Public

    Unbiased offline evaluation on real Kuaishou logs: ranking survives randomized exposure, calibration does not, and no single calibrator fixes both.

    HTML

Repositories

Showing 6 of 6 repositories

People

This organization has no public members. You must be a member to see who’s a part of this organization.

Top languages

Python HTML