Case study · NBFC · Generative AI
How Kogta Financial deployed an AI accountability agent for 5,000 field executives in five weeks with ShellKode FDEs.
Kogta Financial · NBFC
Managers used to chase every missed target. Now Priya does, and she remembers every promise.
5 weeks
From design to deployment, with a two-person FDE team
~5,000
Field executives the agent is designed to support
≥85%
Commitment extraction accuracy target
12
Language models tested against identical business scenarios
The challenge
Month-end accountability ran on managers' time and memory.
At a lender, every month ends in a race to hit disbursement targets. When a field executive falls behind, a manager calls to ask what's holding the files up and when they'll move. Across thousands of executives, no manager team can do that consistently.
A chatbot built on prompts alone wouldn't solve it. The agent had to turn conversations into dated commitments, never disclose targets, never invent loan data, and never end a chat early, every single time. Working out which rules belonged to the model and which belonged in code meant testing real scenarios side by side with Kogta's engineering lead. That called for Forward Deployed Engineers building inside the workflow, not handing over a spec.
Manual follow-ups at scale
Individual follow-ups with ~5,000 executives weren't operationally realistic.
Vague commitments
General assurances couldn't be tracked against pending loan files.
Little visibility into blockers
There was no consistent way to capture delays or spot repeated misses.
Rules an LLM can't be trusted to follow alone
Invalid dates, leaked targets and invented data couldn't be left to chance.
What we delivered
AI for the conversation, code for the rules.
ShellKode FDEs worked directly with Kogta's engineering team to answer two questions about every chat: did Priya get a real commitment, and did she follow every rule getting it?
A conversation that remembers
Priya runs one question at a time, keeping each executive's files and past commitments across sessions.
Chats become accountability data
Commitments, blockers and escalation flags are stored in Amazon DocumentDB, and managers are notified when misses repeat.
Guardrails that don't depend on the prompt
Server-side checks reject vague or invalid dates, block confidential disclosures and correct made-up details.
Free to switch models
Twelve models were tested on the same scenarios, and LiteLLM routing over Amazon Bedrock lets Kogta switch models without changing the app.
The outcome
Every executive gets the same follow-up, and every promise gets tracked.
≥85% extraction accuracy target
Vague promises become specific, validated dates and times.
100% persistence target
Every message and commitment is stored, with nothing lost on reconnects.
15+ guardrail fixes
Found through scenario testing before Priya ever reached an executive.
Managers focused on exceptions
Escalation flags bring managers in only when repeated misses need them.
The stack
What it runs on.
- Models
- Amazon Bedrock, 12 models tested
- Model routing
- LiteLLM over Amazon Bedrock
- Data
- Amazon DocumentDB
What's next
Priya is live. Next comes making every follow-up count.
With conversation, rules and memory separated, refining Priya's behaviour is a configuration change, not a rebuild.
- 01
Rolling out to the full field force of ~5,000 executives
- 02
Re-testing models on quality, latency and cost as new ones ship
- 03
Turning blocker data into insights on why files stall
Talk to us
Bring one conversation your managers repeat. Leave knowing which parts an agent can hold.
30 minutes on one workflow, mapped to the agent, guardrails and model choices behind it. No slides.