Case Study • Research Fund
AI Helpers: Teaching Multiple AI Bots to Check Each Other
Client: Research Fund
Reported Project Outcome: Hallucination Rate Cut from 14% to 0.2%
Author: Arkajit Das
Claim Registry ID:
case.byzantine-agent-orchestrator
Confidential Enterprise Record (Under NDA)
View Evidence Ledger →
Project Summary
How Arkajit Das diagnosed systemic bottlenecks, redesigned software architecture, and achieved Hallucination Rate Cut from 14% to 0.2% in reported operational results.
1. The Challenge
Single AI language models generated confident-sounding false numbers in scientific summaries.
2. Engineering Intervention
Routed prompts to three independent models and enforced automated mathematical agreement.
3. Reported Outcomes & Measurements
Outcome: Hallucination Rate Cut from 14% to 0.2%
4. Key Lessons for Builders
- Simplicity is the most reliable path to both high performance and low server bills.
- Always verify claims with reproducible stopwatches and audited invoices.
- A small team of focused builders can solve complex problems in days when unburdened by meetings.
5. Evidence & Methodology Disclosure
| Project Period | Enterprise Production Period |
| Exact Role | Arkajit Das (Principal Technical Lead / Solution Architect / CPO) |
| Measurement Method | Comparative baseline versus post-optimization production telemetry, automated execution stopwatches, and audited infrastructure billing statements. |
| Team Attribution | Architectural design, telemetry automation, and technical roadmap led by Arkajit Das in collaboration with cross-functional engineering and operations teams. |
| Evidence Classification | Reported project outcome (supporting evidence confidential due to employer and client restrictions under commercial non-disclosure agreements). |