Vishisht
Mishra

Open-source AI safety middleware for LLM chat platforms. It intercepts self-harm and criminal-intent messages before the model sees them, alerts operators and responds with care.

• ·A 3-stage classification cascade (heuristics → semantic embeddings → reasoning LLM); clear-cut cases exit early, so latency and cost stay proportional to risk.
• ·Trained a custom model; the framework benchmarks at 92% recall on SimpleSafetyTests with a 1.2% false-positive rate on XSTest.
• ·MCP server with bearer auth, swappable storage (SQLite, PostgreSQL, Redis) and a REST admin API with webhook alerts. 8.52K+ installs on PyPI.

Self-healing GKE infrastructure provisioned with Terraform and deployed through a zero-touch GitHub Actions pipeline, released as an open-source template.

• ·Every image passes a Trivy vulnerability scan before it ships; releases need no manual steps.
• ·Resilience validated with k6 load tests that trigger HPA (2–10 pods) and cluster autoscaling (1–3 nodes); Prometheus and Grafana monitoring and optimized containers come built in.

A hybrid GNN + Autoformer model forecasting GNSS satellite ephemeris and clock errors, built for an ISRO problem statement. It placed 3rd at the Gateway to Smart India Hackathon.

• ·Multi-stage transfer learning handles extreme data scarcity: the real training set spans just 7 days.
• ·Augmented the data 100× with a fine-tuned Large Time-Series Model, reaching near-Gaussian residuals (Shapiro-Wilk p > 0.05).
• ·Later extended into research applying PRESTO to 5G non-terrestrial networks.

A fleet and ERP platform covering inventory, manufacturing, order tracking and finance, built end to end in Next.js 15 and TypeScript as the only engineer on the project.

• ·Designed the backend and MongoDB data models, turning client operational requirements into workflow systems.
• ·NextAuth authentication and Vitest tests; ReactFlow components visualize operational node graphs.
• ·Added RAG-based stock querying on top of the inventory data.