What Went Wrong with Forward: A Postmortem on AI-Powered CarePods

---

In November 2024, Forward, a healthcare tech startup once backed by over $650 million in funding, abruptly shut down. This is not just a teardown; it’s a lesson in how data systems can make or break a technically ambitious product.

Context: What Forward Tried to Do

Forward aimed to disrupt primary care through its flagship product: The CarePod. These were standalone, AI-powered booths designed for self-administered health checkups. Think of a futuristic kiosk where you could measure vitals, run diagnostics, and receive AI-generated health insights—all without a doctor present.

Sounds like something out of sci-fi. Unfortunately, the execution never matched the vision.

---

1. Real-Time Monitoring: Or Lack Thereof

One of the biggest red flags was the absence of robust real-time monitoring. With hardware deployed in the wild, capturing telemetry in real-time should be table stakes. Yet many pods were reporting inconsistent data, and issues like malfunctioning blood draw mechanisms went unnoticed for hours, sometimes days.

Technical Failure:

• No Kafka or Pulsar backbone for telemetry.

• Minimal use of stream processing (e.g., Apache Flink or Spark Structured Streaming).

• Delayed ingestion into observability dashboards.

Solution:

Set up a Kafka pipeline to ingest pod telemetry in real time, backed by ksqlDB or Flink for on-the-fly analysis. Use Prometheus/Grafana for alerting. This alone could have prevented a huge chunk of user dissatisfaction.

---

2. Data Integration: Fragmented and Brittle

Each CarePod generated data from sensors, user inputs, and internal AI predictions. But these data streams remained siloed. This made it near impossible to generate a comprehensive 360-degree view of usage or system health.

Technical Failure:

• No data lake or lakehouse for unifying structured and semi-structured data.

• Poor schema evolution strategy.

• Lack of batch + streaming reconciliation.

Solution:

Implement a lakehouse architecture using Delta Lake or Apache Iceberg on S3/GCS. Use schema registries and enforce metadata lineage using tools like Amundsen or DataHub.

---

3. Predictive Maintenance: Totally Absent

Predictive maintenance could have been a game changer. With enough historical sensor data, basic ML models could predict failures before they happened. Instead, they relied on users reporting issues manually—if at all.

Technical Failure:

• No ML pipeline for predictive analytics.

• Inconsistent logging of component-level sensor metrics.

• No use of time-series modeling or anomaly detection.

Solution:

Use AWS SageMaker Pipelines or Airflow + Vertex AI to orchestrate ML workflows. Implement time-series models with Prophet or LSTM, feeding into anomaly detection dashboards.

---

4. Feedback Loops: No Voice of the User

Users interacting with the pods had limited ability to give feedback. Worse, the feedback that was given was not systematically captured or analyzed.

Technical Failure:

• No in-pod feedback capture UI.

• No sentiment analysis on support tickets.

• No NLP pipeline to categorize qualitative user feedback.

Solution:

Incorporate UI/UX event tracking via Segment or Rudderstack. Feed textual feedback into a simple NLP pipeline using spaCy or HuggingFace to cluster and analyze sentiment.

---

5. Scalability: Poorly Designed Infra for Growth

As more pods were deployed, the infrastructure started to buckle. ETL jobs failed more often. Dashboards lagged. Data quality degraded. It was clear the backend was not designed for scale.

Technical Failure:

• Monolithic ETL jobs (likely Python scripts + cron).

• No DAG orchestration via Airflow or Dagster.

• Weak data validation.

Solution:

Migrate to modular DAG-based orchestration. Adopt dbt for transformation and Great Expectations or Soda for data quality checks. Use BigQuery/Snowflake with partitioning and clustering for performance at scale.

---

Closing Thoughts

Forward's failure is often framed as a product-market fit issue. But from where I sit, it’s a data infrastructure issue. The product couldn’t adapt to user behavior because it wasn’t instrumented to understand it. It couldn’t scale operationally because the systems weren’t built to handle growth. And it couldn’t iterate because the feedback loop was broken.

If you’re working on a health tech product—or any hardware-based service—the takeaway is simple: invest in your data stack early. It’s not a cost center. It’s the foundation for everything you hope to build.

---