Bloomit Agent Platform
AI work agents for CRM — retrieve knowledge, call tools, update records, escalate to humans.
Problem
Ops teams need AI that completes work across systems, not a chatbot that only answers questions.
Overview
Bloomit Labs is a live product that deploys AI agents for operational processes. Agents score work, update systems, fire follow-ups, and ping a person only when rules require judgment.
Solution
An agent loop with retrieval, tool calling, write-backs, and explicit human checkpoints — wired into tools teams already use.
Architecture
Client → API → auth → retrieval over knowledge → LLM with tools → background work for long steps → CRM/inbox APIs → audit + human approval path.
- Request
- Auth
- Retrieve
- Agent + tools
- Jobs
- CRM / APIs
- Human gate
Implementation
- Tool/function calling for CRM and inbox actions.
- RAG-style knowledge use for grounded responses.
- Process-shaped workflows (intake → action → follow-up).
- Product UI and narrative shipped on Next.js.
API design
Workflow-oriented endpoints for starting agent runs, inspecting status, and collecting human decisions.
Tradeoffs
- Hosted LLM APIs for speed vs self-hosted models for control.
- One high-ROI process first vs automating every team at once.
Challenges
- Keeping tool calls scoped and reversible where possible.
- Deciding escalation rules that ops can live with.
- Grounding answers in retrieved sources.
Reliability
- Retries on transient provider/API failures.
- Human gate for irreversible or high-risk actions.
Security
- Authenticated product surfaces.
- Least-privilege access to connected systems.
Performance
No public load numbers published. Latency and token cost are tracked per run during development and pilots.
What I learned
- Agents fail in production when tool contracts and escalation rules are vague.
- Shipping into existing CRM/inbox beats a parallel dashboard nobody opens.
Highlights
- Human approval when judgment is required
- Writes into existing CRM / inbox / docs
- Public product at bloomitlabs.com