Deployed systems have a knack for humbling even experienced engineers. You trust your stack—until one morning it silently fails. Last week, during DheeGPT’s development, that exact scenario unfolded and forced us to rethink our infrastructure.
The Day Cassandra Let Us Down
We chose Apache Cassandra to power DheeGPT, our agentic AI engine for Indic languages. It had handled thousands of writes per second in our legacy and production systems — however in the development environment of our newest platform, a routine query returned zero rows:
- The data was unquestionably in the table.
- A secondary-index query returned no results.
- A full table scan confirmed the rows existed—but the index was stale.
No errors. No alerts. Just missing data. Numerous minutes of debugging later, we traced the fault to Cassandra’s secondary index divergence. Invisible failures like this are unacceptable for real-time AI and and automations which are mission critical.
Where Cassandra’s Design Reveals Strain
Cassandra excels as a write-optimized ledger. Yet when you layer on dynamic queries and tight consistency, its limits appear:
- Secondary indexes can silently desynchronize.
- Rigid schemas force costly refactors if access patterns evolve.
- JVM GC pauses inject unpredictable P99 spikes.
As we built DheeGPT, these factors made us question whether Cassandra could sustain our reliability bar.
The Pivot to ScyllaDB During Development
Facing looming deadlines, we swapped in ScyllaDB mid-development. The transition was surprisingly smooth:
- Zero code changes — Same CQL, same Spring Data drivers.
- C++ Seastar lock-free, thread-per-core design.
- Up to 10× throughput on identical hardware.
- Sub-10 ms P99 latencies, even under traffic surges.
Though ScyllaDB launched in 2016, it still flies under the radar. License concerns and slower early evangelism held it back—yet today it’s a mature, high-performance alternative.
How DheeYantra Leverages Robust Infra
At DheeYantra, we build Dhee.AI Digital Employees that:
- Power 24×7 customer support in Hindi, Tamil, Bengali, and more
- Automate document processing and data extraction
- Orchestrate workflows across enterprise systems in real time
For us, infrastructure reliability isn’t optional—it’s the core of our product promise. With ScyllaDB, we maintain sub-second response times at scale, reduce cluster size (and costs), and shift focus from firefighting infra to advancing AI logic.
Key Takeaways
- Revisit your defaults. Even proven tech can hide silent failure modes.
- Decouple API from engine. A compatible interface lets you switch back ends with minimal friction.
- Benchmark real workloads. Surface tail-latency and consistency issues before they reach prod.
- Invest in observability. Detect index divergence proactively, not after customers complain.
What’s Next: Digital Employees Powered by Indic Agentic LLMs
We’re excited to announce that next week we’re rolling out the new DheeGPT platform alongside our first Indic Agentic LLMs. These models will underpin Dhee.AI Digital Employees—reliable, multilingual agents tailored for Indian enterprises.
👉 Keep an eye on DheeYantra’s page for the official launch and demos.
👉 Reach out to explore how Dhee-powered Digital Employees can transform your workflows—whether in support, finance, HR, or field services.
Reliable AI starts with dependable infrastructure—and yes intelligent digital employees (we gave got you covered there).
Let’s build the future, together.

Sreekumar (KJ) has been a hobby programmer from school days. Codemarvels is his personal blog from the year 2010, where he writes about technology, philosophy, society and a bit about physics.
He now runs a conversational AI company – DheeYantra – focusing his efforts to help businesses improve operational efficiency using digital employees powered by AI.
Leave a Reply