JUNE 2026 – PRESENT
INTELLIGENT CRIME DATABASE CONVERSATIONAL AI
AI-Powered Multilingual RAG Search & Predictive Analytics
01 // PROJECT_SUMMARY
Led AI development by building LLM-powered conversational search, NLP intent classification, Retrieval-Augmented Generation (RAG) pipelines, and predictive analytics for multilingual crime intelligence and investigation support for KSP (Karnataka State Police) Crime Database.
PythonNLTKTableauLLMsRAGFAISSNLP
02 // STRIDE_THREAT_MODELING_LOGS
| THREAT_CATEGORY | EXPLOIT_VECTOR | MITIGATION_STRATEGY |
|---|---|---|
| Prompt Injection (AI Specific) | Malicious users input adversarial prompts to force the LLM to output classified investigation case details. | Implemented LLM Guardrails (input-output validation) and system-level prompt isolation. |
| Data Leakage via Embeddings | Access to the FAISS vector database allows attackers to reconstruct original case documents. | Strict role-based access control (RBAC) on the API endpoints serving the embedding query. |
| Denial of Service (DoS) via Token Consumption | Attackers submit extremely long queries to deplete API tokens and throttle system resources. | Rate limiting by user session and prompt length token limits at the API gateway. |
03 // ARCHITECTURAL_SANDBOX_SCHEMAS
FILE_DUMP // DIAGRAM_NODES.LOG
- Data Pipeline: CSV/JSON crime log ingest parsed, cleaned, and normalized with NLTK.
- Embedding Vector Database: Documents chunked and stored in a local FAISS database using text-embedding models.
- RAG Core: Query-context retrieval matching case IDs combined with LLM prompting for natural responses.
- Analytics Dashboard: Integrated Tableau dashboards displaying predictive hot-spots and monthly crime trends.
04 // ARCHITECTURE_LESSONS
- LLMs are prone to hallucinating penal codes (e.g., IPC/BNS sections). Grounding the model strictly to verified datasets is mandatory.
- Handling multi-lingual search queries requires semantic embeddings that map local dialects (e.g., Kannada descriptions) to centralized crime schemas.
- Performance optimization is crucial when querying massive crime databases with low latency.
05 // TARGET_OUTCOMES
- Optimized crime case search times, transforming keyword-based filtering into semantic conversations.
- Presented predictive insights on crime hotspots using Tableau visualizations.