Mihir Srivastava projects

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

• ai
• devtools
• infrastructure
• Python
• TypeScript
• MCP

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

• systems
• devtools
• security
• Go
• PostgreSQL
• MCP

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

• ai
• fullstack
• Python
• FastAPI
• TypeScript
• React

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

• ai
• research
• rl
• Python
• Gemini API
• AIME 2025

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

• ai
• Python
• Groq
• Gemini
• LM Studio

Modernized Cisco ASA LiNA LDAP authentication by extracting a Linux agent; moved to upstream OpenLDAP, closed 45+ CVEs, and made authentication 5x faster.

• systems
• security
• devops
• C
• Linux
• OpenLDAP

Globally distributed predictive file cache built in Go, trained on 100K+ access logs and achieving 95%+ prediction accuracy.

• systems
• devtools
• Go

ERSP research using PyRTL and Egglog to recover fold abstractions from synthesized netlists and verify Boolean simplifications.

• research
• systems
• Python
• Boolean Logic
• PyRTL
• Egglog

Python CLI that converts natural-language intent into shell commands with Groq-backed generation, syntax validation, and configurable execution guardrails.

• ai
• devtools
• Python
• GroqCloud

Facial-emotion classifier trained on 35K+ images with 87% accuracy, paired with Gemini 1.5 Pro mood recommendations; first place among 53 teams.

• ml
• cv
• PyTorch
• OpenCV
• Gemini
• Streamlit

Expo/React Native prototype for discovering and reporting UCSB campus utility conditions; winner of Best UI/UX among 10 teams.

• mobile
• React Native
• Expo
• React Navigation
• twrnc

Compared five classifiers for red and white wine quality, reaching 84% accuracy from 11 physicochemical features and placing top five at the UCSB Datathon.

• ml
• Scikit-learn
• Pandas
• Jupyter Notebook

Research study tracing Moore’s Law from 1965 through heterogeneous computing and its continuing influence on machine-learning hardware; Polygence Symposium finalist.

• research