Prasad Kavuri

Real retrieval-augmented generation with Transformers.js embeddings and ChromaDB — runs entirely in your browser.

Shows how enterprise knowledge can be retrieved with source traceability, relevance controls, and citation — not hallucination.

Real multi-model routing across Llama 3.1 8B, 70B, and Mixtral — see live latency, cost, and quality trade-offs.

Routes each AI request to the optimal model — from fast WASM models for simple tasks to Qwen3.6-27B for agentic reasoning — demonstrating LLM FinOps and decision intelligence at the platform level.

Semantic search with real sentence-BERT embeddings and UMAP visualisation of the embedding space.

High-throughput semantic search over large corpora with real-time filtering — the retrieval layer for any serious AI product.

Closed-loop LLM evaluation pipeline — semantic fidelity, hallucination detection, guardrails, and CI gating in action. Production-derived eval thresholds — calibrated from real Krutrim deployment patterns.

Catches quality regressions before they reach production — the governance layer that separates AI experiments from AI platforms.

A reusable Chrome AI Skill that audits webpage accessibility using on-device Gemini Nano.

On-device AI inference with zero server dependency — the architecture pattern for compliance-sensitive enterprise tooling.

CrewAI-powered agents with real LLM calls via Groq — Analyzer, Researcher, and Strategist collaborating in real time.

Demonstrates governed agentic workflows with human-in-the-loop approval checkpoints, audit trails, and role-based orchestration — safe for enterprise deployment.

Model Context Protocol in action — watch an LLM discover and call tools to answer questions about Prasad's background.

Shows how standardized tool protocols reduce integration overhead and make agent capabilities composable across enterprise systems.

Org-wide AI governance dashboard — RBAC, group spend limits with token-cost tracking, and structured observability feed.

Operational guardrails for enterprise AI: RBAC, spend analytics, token budgets, and structured observability in a single control surface.

Three-tier privacy-first AI pipeline: BERT NER redacts PII in the browser via Transformers.js ONNX, a HITL gate governs the handoff, and Groq produces an executive summary from the sanitized payload only.

Demonstrates the governance-first agentic handoff pattern enterprises need for regulated AI workflows: edge extraction, explicit HITL approval, and auditable cloud orchestration.

Live auth.md protocol implementation — AI agents register anonymously, claim with email + OTP, then call MCP tools with a Bearer credential.

Demonstrates agent-native authentication: anonymous registration, email claim, OTP upgrade, and authenticated tool access — the identity primitive every agentic platform needs.

Perception → reconstruction → agent reasoning. Precomputed 3D mesh playback with drift correction visualization and LLM spatial query layer. Controllable parametric spatial design — refine generated scenes with natural-language instructions. Changes are validated, diffed, and auditable.

Brings LLM reasoning into spatial and operational planning — policy-aware world models that are auditable, diffable, and simulation-ready.

Streaming full-context assistant over my experience with optional retrieval-enhanced grounding and cited context cues.

Demonstrates conversational AI grounded in structured knowledge — RAG + LLM working together on a real corpus.

Paste a job description — get multi-dimension fit scoring, HITL-gated tailoring, and an ATS-optimized resume with drift detection.

AI-powered resume tailoring at scale — shows LLM orchestration applied to a high-frequency, measurable business workflow.

Florence-2 image captioning and OCR running in-browser via Transformers.js — no server, no API key.

Edge-deployed vision AI with zero server cost — the architecture pattern for privacy-sensitive document and image workflows.

Live ONNX benchmark comparing INT8 vs FP32 inference — real file sizes, real latency, real quality diff.

Demonstrates 4-bit MoE quantization delivering 70%+ memory reduction with minimal quality loss — the cost lever most teams overlook.