Solutions

BlueKyte’s AI Offerings

• Use-case
• Solution / Approach*
• USPs
• Time to PoC*
• Needed for PoC
• Remarks / Questions

• 1
  AI-Powered Hybrid Search
• • Fine-tuned BERT
  • Similarity search
• • Context
  • Ambiguity Resolution
  • Auto-Complete
• 2-4 weeks
• • A db of records
  • Architecture, schema, etc.
• • What is the tech-stack for the existing search?

• 2
  Data Interrogation & Visualisation
• • LlamaIndex
  • Sagemaker (or on-premise)
  • ELK Stack
• • Set automated activity alerts in NL
  • Volume and trend analysis
  • Map visualisation
• 4-6 weeks
• • Mastered prod data
  • Schema
  • Types of desired data interrogation
• • What are the client’s data interrogation needs?
  • What is the expected scale?

• 3
  Key Information Extraction
• • SOTA models + light LLMs
  • Azure ML
• • Upwards of 98% accuracy on CORDS & FUNSD datasets
  • TPP by default
• 4-6 weeks
• • Existing benchmarks
  • All possible templates. and a sample dataset for testing
  • Supervision
• • More information on the nature of documents, templates, and problems with current model.
  • Validation and confidence score calibration flows.

• 4
  Automated document sanity checks at scale
• • Toolformer
  • Scalable vector stores
  • MoEs
• • Orchestration
  • AI Agents
  • Plots & Custom Dashboards
  • Reliability
  • Security
• 6-12 weeks
• • Protocols, rulebooks, etc.
  • Validation benchmarks
  • Comprehensiveset of undesirable behaviours
• • Financial data retrieval - schema, db, etc.
  • What other tools have been considered?

• 5
  Role-specific Chatbot
• • Custom MoFEs
  • LangChain or LlamaIndex
  • Vector store
  • Proprietary modules for maintaining state, recording and accessing sessions and RLHF
• • Scalable Dataset Inferencing & Q/A
  • SOTA prompt operators packaged as buttons / query auto-complete
  • Usage pattern recognition and personalised workflow suggestions
• 4-8 weeks
• • Sample customer data - structured and unstructured
  • Classes of expected questions
  • User behavior metrics and logs (db and schema)
  • Custom workflow wishlist and existing system architecture
• • Expected user roles (int + ext)
  • Constraints
  • What is the desired speed of response (tokens per sec)?
  • How many tiers of custom and personalised actions/workflows?

• 6
  Fraud Detection & Prevention
• • Combination of rule-based and ML-based algos
  • DL, if high risk
• • Use GenAI for alerts, reduced customer friction and incident reporting
• 6-12 weeks
• • Financial data and a classification of instances of fraud, along with priority
  • Internal processes and regulation
• • How much data is generated per day?
  • What are the priorities wrt batch and stream processing?
  • What are the systems currently in place to detect/prevent fraud?

• 7
  Compliance Reports & Alerts
• • Automated web scraping for relevant updates
  • Custom MoEs
  • Deep Learning
• • Natural language alerts
  • Hidden insights from prod data
• 4-8 weeks
• • Protocols, latest regulatory material and rulebooks
  • Real time prod data
• • Info on static and dynamic laws/ regulations/protocols
  • What factors to take into account?