Vishisht
Mishra

Humane Proxy

April 2026

HumaneProxy is a production-grade, privacy-first AI safety middleware designed to protect humans—not just models. Operating as a plug-and-play reverse proxy and official Anthropic MCP server, it intercepts self-harm and criminal intent using a highly optimized 3-stage cascading inference pipeline (Heuristics → Semantic Embeddings → Reasoning LLMs) before the prompt ever reaches the upstream LLM.

Engineered for enterprise scalability, it features exponential time-decay risk tracking, distributed storage backends, and native integrations for LangChain, LlamaIndex, etc. Beyond real-time crisis intervention, it ships with an automated DevSecOps benchmarking GitHub Action, empowering teams to enforce safety guardrails in their CI/CD pipelines with zero underlying code changes.

GKE Pipeline

February 2026

Architected and provisioned a self-healing Google Kubernetes Engine (GKE) infrastructure utilizing Terraform as Infrastructure-as-Code (IaC), establishing a highly available and robust foundation for scalable microservices.

Engineered a secure, zero-touch CI/CD pipeline driven by GitHub Actions, natively integrating Trivy-based vulnerability scanning to ensure that only rigorously tested and fully compliant container images are automatically deployed.

Validated system elasticity and resilience through comprehensive k6 load testing to dynamically trigger Horizontal Pod Autoscaling (2–10 pods) and Node Auto-provisioning (1–3 nodes), subsequently integrating Prometheus and Grafana for real-time observability and publishing the complete architecture as an open-source template.

Fleetware

October 2025 - January 2026

Built from scratch, Fleetware is a full-stack ERP system designed for fleet and warehouse management, handling end-to-end operations like inventory control, purchase & sales orders, vendor/customer management, manufacturing workflows, and financial tracking. It features a complete real-world business logic engine with automated stock updates, order lifecycle handling, and scalable data architecture.

Entirely designed and developed independently, focusing on complex system design, clean architecture, and production-level business workflows.

A previous version of the prototype is currently available as demo.

PRESTO — Satellite Error Forecasting

October 2025

Engineered a novel hybrid deep learning framework integrating Graph Attention Networks (GATConv) and Autoformers to forecast GEO/MEO satellite ephemeris and clock errors for an ISRO problem statement.

Mitigated extreme data scarcity (only 7 days of real data) by designing a 3-stage transfer learning pipeline, utilizing a fine-tuned Large Time-Series Model to generate a 100x synthetic dataset for robust pre-training.

Optimized probabilistic forecasting accuracy using Negative Log-Likelihood and semi-parametric residual isolation, successfully yielding statistically rigorous, near-Gaussian error distributions (Shapiro-Wilk p > 0.05).