Job description:
AI Engineer
===========
**Location:** SF Bay Area ( 4
- days in office )
**Experience Level:** 3–5
- years
**Stack:** Python, PyTorch, ML, LLMs, Django
**Type:** Full-time
* * *
### **🧠 About Coris**
[Coris](https://www.coris.ai/) is building the AI-first trust layer for global commerce. We partner with leading platforms, marketplaces, payment providers, and banks to transform how small business onboarding, monitoring, and lifecycle decisions are made
- using AI on the ground to drive faster, smarter actions with less friction.
One of our customers described us as **Cursor
- Lovable for risk teams**: flagging bad actors, assisting in investigations, and autonomously acting to mitigate fraud losses in real time.
Backed by top-tier investors and founded by experts in the payments domain, Coris is reimagining how risk gets done
- not with legacy rule engines, but with **domain-specific AI that thinks like your best risk analyst at scale.** We help customers scale their expertise, move faster, and unlock growth
- without compromising safety.
* * *
🚀 **Why this role matters**
----------------------------
Fraud detection and Risk mitigation is a uniquely hard ML problem:
* **Adaptive adversaries**
- fraudsters continuously evolve tactics, so models must adapt faster than static rules.
* **Data sparsity and imbalance**
- only a tiny fraction of transactions are fraudulent, but they cost millions.
* **Latency and scale**
- decisions need to happen in **tens of milliseconds** at **hundreds of millions of events per month**, without ballooning infra costs.
This role is for someone who wants to **optimize language models for fraud/risk contexts** _and_ build the **backend infra** that productionizes them at scale.
* * *
🥷 **What you’ll do**
---------------------
### **AI/ML (~50%)**
* Fine-tune, distill, and quantize **LLMs and small language models (SLMs)** for fraud detection tasks: entity resolution, anomaly detection, customer communication classification, synthetic data generation.
* Optimize inference so our models run **fast and cost-efficiently** in production
- using techniques like **lightweight fine-tuning (LoRA/PEFT)**, **quantization to smaller precisions**, and modern serving frameworks (e.g. **vLLM, TensorRT**)
* Build training/eval pipelines for fraud models that balance **recall** (catch fraud) with **precision** (minimize false positives).
* Create golden datasets, adversarial test sets, and online/offline evaluation harnesses that mirror **real-world fraud evolution**.
* Build feature engineering pipelines extracting various signals including the non-obvious latent ones.
### **Backend (~50%)**
* Architect and own **Python/Django services** that integrate model predictions directly into customer-facing APIs.
* Model complex fraud/risk data in **Postgres**; ensure queries and aggregations scale to billions of records.
* Build/Operate/Enhance data ingestion pipelines from **Stripe, Adyen, and other payment processors**, handling near real-time volume.
* Ensure observability with logs, metrics, and drift detection to catch when fraud tactics change.
* * *
✅ **You may be a fit if you have**
----------------------------------
* 3
- years building production systems in **Python/Django** with **Postgres**.
* Hands-on experience fine-tuning and optimizing **LLMs/SLMs**, ideally in **fraud, anomaly detection, or adversarial domains**.
* A track record of reducing **latency/cost** in ML inference without compromising accuracy.
* Comfort working across the stack
- from PyTorch profiling to Django APIs.
* An experimental but practical mindset: ship fast, measure rigorously, iterate.
* * *
🙏 **Nice to have**
-------------------
* Prior work with **imbalanced datasets** (e.g., 1 in 10,000 fraud cases).
* Knowledge of **feature stores, online learning, and temporal aggregation** for fraud models.
* Familiarity with **regulatory requirements** around PII, KYC/AML, and compliance in financial data.
* * *
🚀 **Success in 3-6 months**
----------------------------
* A distilled/quantized fraud model running in prod with **2-3x lower latency/cost** than baseline, catching more fraud with fewer false positives.
* A **robust pipeline** for fine-tuning/evaluating fraud models that the team trusts.
* Django services powering **real-time fraud scoring APIs** integrated with Stripe/Adyen data flows.
* * *
🤝 **How we work**
------------------
* Bias toward action, measurable impact, and **staying ahead of adversaries**.
* Everyone owns their code in prod
- from training to inference to APIs.
* Fast iterations with real customer feedback; clear metrics drive decisions.
* In-person culture with at least 4 days a week in our Palo Alto Office.
* Like any other high growth startup, we go much beyond the usual 40/50 hrs per week.
* We need high energy, high agency individuals who go the extra mile to get things done.
* * *
💰 **Compensation**
-------------------
Competitive salary
- equity
- benefits.