AI agents that act, safely.
An AI agent is software that plans, uses tools and checks its own work to finish a task. Sandeep Kavety designs and ships governed agentic systems for organisations: orchestration, integrations, evaluation and the guardrails that make them safe to run on real data.
- 5
- production agentic platforms
- 30 days
- CFO analytics to production
- 29
- agents in one content studio
What I build
01
Plan, execute, reflect
Multi-agent orchestration with a planning step, tool execution and a reflection pass before anything is returned.
02
Capability broker and approvals
Agents only get the capabilities they need, and actions are approved by risk tier before they run.
03
No secret reaches a model
Credential handles and a taint engine keep keys and sensitive values out of prompts.
04
MCP and integrations
Model Context Protocol and API integrations so agents work inside the systems you already have.
05
Evaluation and injection defence
Agent evaluation suites and prompt-injection defences, measured before launch, not after.
06
Retrieval over your data
RAG with pgvector, OpenSearch or Kendra, with per-tenant isolation for multi-customer products.
Evidence
Insighter
Governed multi-agent automation built in Python, FastAPI and DuckDB. Shortlisted for IIT Mandi HIVE 3.0 finals.
See more →
Becide.ai
A multi-tenant CFO analytics platform with a Bedrock ReAct orchestrator, live in 30 days and built to Australian privacy rules.
See more →
Working papers
Two papers on conversational decision intelligence and the one-agent-many-hats architecture.
See more →
MCP server design
A Model Context Protocol server where a customer's AI assistant can read, draft and schedule through scoped tools, with human confirm links for anything that goes out.
See more →
AI agents: common questions
What is an agentic AI system?
An agentic system uses a language model to plan steps, call tools or APIs, and verify results, rather than only answering a prompt. It needs orchestration, permissions and evaluation to be reliable.
How do you keep AI agents safe in production?
Capabilities are brokered rather than granted broadly, actions are approved by risk tier, secrets are handled by reference so they never reach the model, and behaviour is evaluated before release.
Do you build a fleet of agents or one agent?
It depends on the problem. In Insighter, a single reasoning engine adopts distinct roles through skills, which avoided the coordination overhead of a planner, critic and coder passing messages.
How long does a first agent take?
A thin production slice is usually weeks, not quarters. The CFO analytics platform for an Australian client reached production in 30 days.
Related