Mihir Srivastava

profile

Researches persona-based LLM debate at UCSB, modernized Cisco ASA LDAP authentication by closing 45+ CVEs and reducing latency 5x, and built predictive caching at Spillbox with 95%+ accuracy over 100K+ logs.

current focus

• Systems and security work around auth paths, network appliances, and infrastructure correctness.
• Machine learning and NLP research, including self-questioning language models and tool-verified evaluation.
• Developer tools and agent workflows that make local build, research, and command-line work easier to inspect.

experience

• Software Engineering Intern / EverpureJune 2026 – Sept 2026 / Santa Clara, CA

  Built an infrastructure metrics MCP server for an internal AI triage agent to more accurately differentiate infrastructure issues from FlashBlade simulator code issues, reducing false positives by 89% and increasing accuracy by 45%.

• Undergraduate Researcher / UCSB NLP GroupSept 2025 – Present / Santa Barbara, CA

  Researching persona-based LLM debate by designing a task-adaptive persona generator and evaluation pipeline to study how reasoning-style diversity influences debate outcomes, convergence, and failure modes on Humanity's Last Exam-style problems.

• Software Engineering Intern / CiscoJune 2025 – Sept 2025 / Boxborough, MA

  Modernized Cisco's ASA LiNA LDAP authentication subsystem by extracting it into a standalone Linux agent, replacing a heavily customized fork with upstream OpenLDAP, closing 45+ CVEs, and reducing latency by 5x.

• Software Engineering Intern / SpillboxJuly 2024 – Oct 2024 / Remote

  Engineered a high-performance predictive file caching system with distributed network architecture, processing 100K+ log files with 95%+ accuracy.

• Software Engineering Intern / OpenROAD Project | University of California, San DiegoJune 2022 – Oct 2022 / University of California, San Diego

  Improved DevOps efficiency by modernizing the Jenkins pipeline-interface - added features such as version tracking & performance comparison, optimized workflow & redesigned user interface for test case monitoring & development.

selected projects

• AI Infrastructure Triage AgentJune 2026 – Sept 2026

  Built infrastructure tooling and a custom agent harness for an internal AI triage agent, reducing false positives by 89% and increasing accuracy by 45%.

• Agent MailboxFeb 2026

  Secure PostgreSQL-backed MCP message bus for Codex and Claude Code agents, with direct and broadcast delivery, priority long polling, acknowledgements, and auditable events.

• Policy RadarDec 2025 – Jan 2026

  Full-stack federal policy research agent that searches nine official government sources and streams source-backed answers with full-text and PDF retrieval.

• Persona Based Large Language Model DebateSept 2025 – Present

  Evaluation harness comparing single-shot, majority-vote, and persona-conditioned LLM debate on AIME 2025, GPQA, and HLE-Verified.

• LatticeAug 2025 – Present

  Local agentic router that runs multi-step builds and tests in isolated workspaces, records structured logs, and packages only fresh deliverables.

recognition

• 1st place (FacEmotion) / Data Science SB: Won 1st place among 53 teams; 87% accuracy; 35K+ images
• Best UI/UX (Utility Unraveled) / UCSB: Won best UI/UX among 10 teams
• Polygence Symposium finalist / Polygence: Moore's Law: Its Impact Past, Present & Future
• Logistics Director (SBHacks) / SBHacks: Organized UCSB’s largest hackathon with 300+ attendees; raised $30k+

stack