wehub-resource-sync
41 lines
1.7 KiB
Markdown
41 lines
1.7 KiB
Markdown
# Article Digest -- Proof Points
|
|
|
|
Compact proof points from portfolio projects. Read by career-ops at evaluation time.
|
|
|
|
---
|
|
|
|
## FraudShield -- Real-Time Fraud Detection
|
|
|
|
**Hero metrics:** 99.7% precision, 50ms p99 latency, $2M/year fraud prevented
|
|
|
|
**Architecture:** Kafka Streams ingestion → real-time feature computation (200+ features, sliding windows) → ensemble model (XGBoost + neural network) → decision engine with configurable thresholds → human review queue for edge cases
|
|
|
|
**Key decisions:**
|
|
- Chose streaming over batch to catch fraud in real-time (batch had 4-hour delay)
|
|
- Ensemble approach: XGBoost for speed + neural net for complex patterns
|
|
- Built custom feature store for real-time features (Redis-backed, 5ms reads)
|
|
|
|
**Proof points:**
|
|
- Reduced false positives 60% vs previous rule-based system
|
|
- Handles 10K transactions/second peak load
|
|
- 500+ GitHub stars, adopted by 3 fintech startups
|
|
- Conference talk: "Real-Time ML at Scale" (MLConf 2023)
|
|
|
|
---
|
|
|
|
## LLM Eval Toolkit -- Evaluation Framework
|
|
|
|
**Hero metrics:** 15 built-in metrics, CI/CD integration, used by 200+ developers
|
|
|
|
**Architecture:** Pluggable metric system → test suite runner → regression detection → GitHub Actions integration → Slack alerts on regressions
|
|
|
|
**Key decisions:**
|
|
- Metrics as code: each metric is a Python function with clear interface
|
|
- Deterministic testing: seeded prompts + temperature 0 for reproducible evals
|
|
- Cost tracking: each eval run logs token usage and estimated cost
|
|
|
|
**Proof points:**
|
|
- Caught 3 production regressions before deployment in first month
|
|
- Reduced eval cycle from "vibes check" to structured 15-minute CI run
|
|
- Open source, 200+ weekly active users on PyPI
|