Pratyus Kumar
Mansingh

Problem: Managing multiple agents and passing the right context between them became difficult as workflows grew.

Solution: Built LangGraph orchestration with MCP tools for specific task routing, shared memory, and runtime sub-agent spawning.

Problem: Running tools like Metasploit/Nmap directly against the application created security and isolation concerns.

Solution: Designed Docker-based sandboxed environments to safely execute security tools and collect results.

Problem: Security testing exposed risks of sensitive project information leaking through AI inputs and outputs.

Solution: Added input and output guardrails to control project-specific information at multiple AI layers.

Problem: Large source files and codebases exceeded LLM context limits during vulnerability analysis, causing API errors and incomplete scans.

Solution: Built context-aware chunking with temporary memory to track code relationships and selectively pass relevant chunks to the LLM, preventing context-limit errors during large-codebase analysis.

Problem: Multiple users running large scans overloaded a single FastAPI instance.

Solution: Scaled FastAPI horizontally to 5 replicas to handle concurrent scan workloads reliably.

Problem: High CPU/RAM usage during large scans affected application performance.

Solution: Used load balancing and EC2 auto-scaling to distribute workloads and add capacity as demand increased.