Raava

Raava · agentic systems

We build software where agents do the work.

Raava builds agentic systems: AI agents on real operations, with people in the loop where it counts. A contractor's CRM, a lead pipeline, a Foundry toolchain and the memory behind it.

The loop every project shares

Agents act through narrow, typed tools. Writes are planned before they are applied. A person reviews what the model is not sure of. Every run is traced, so the next one starts from evidence.

The loop: an agent takes the task, acts through typed tools, plans or runs in shadow, a person takes the hard calls, and every run is traced as evidence for the next. Agenttakes the task Typed toolsMCP · CLI · JSON out Plan or shadownothing applied blind A persontakes the hard calls Traceevery run recorded evidence for the next run The loop: an agent takes the task, acts through typed tools, plans or runs in shadow, a person takes the hard calls, and every run is traced as evidence for the next. Agenttakes the task Typed toolsMCP · CLI · JSON out Plan or shadownothing applied blind A persontakes the hard calls Traceevery run recorded evidence for the next run
01

CrewOps

In production
CrewOps: a contractor's AI assistant works the records over MCP with a per-person token, while background agents route leads, link email and read scopes, in shadow until they match people's decisions. A contractor's AI assistant MCP · 9 tools · token scoped to one seat read · add · move CrewOps records leads · claims · jobs · docs Route leadagent Link emailagent Read scopeagent in shadow until they match what people decided

An agent-first CRM for roofing and storm-restoration contractors.

A contractor's AI assistant works the same records as the crew, through an MCP server with a token scoped to one person's seat. Behind the app, small agents route leads, link email and read scopes, in shadow until they match what people decided.

MCP tools
9
Platforms
Web · iOS
Agents
Shadow-first
Read the case study →
02

Qwalo

Live · first client
Qwalo: a seller's reply goes to a fast model, then a stronger model when the first is unsure, then a person. Each settles what it can. Human calls go back to the models as examples. Seller replies Fast modelsettles the clear ones unsure Stronger modela second look still unsure A personmakes the call Settledfollow upor drop each human call returns to the models as an example

Lead qualification run by agents, with a person on the hard calls.

A fast model settles the clear seller replies, a stronger one takes a second look, and a person takes the rest. Every human call goes back to the models as an example. Sending stays off until a backtest clears 90%.

Reply gate
≥ 90%
Cascade
3 tiers
Runs on
Cloudflare
Read the case study →
03

Orthanc

Shipping · v0.5.0

AGENT SESSION · REPLAYorthanc 0.5.0

  1. $ orthanc verify --agent — ok · profile prod · token valid (212 ms)
  2. $ orthanc search "work orders" --agent — 3 resources · 1 object type, 2 datasets (388 ms)
  3. $ orthanc ontology object-type-list ops --agent — 41 object types · WorkOrder found (506 ms)
  4. $ orthanc ontology object-count ops WorkOrder --agent — 1,284 objects (274 ms)
  5. $ orthanc ontology action-validate ops close-work-order '{"workOrder":"WO-1042"}' --agent — valid · not applied (331 ms)

Demo. The commands are real Orthanc 0.5.0. The results are made up.

A command-line tool agents use to build on Palantir Foundry.

People and coding agents drive the same binary. Every command can answer in one JSON envelope, every write prints a plan first, and nothing changes in Foundry without --apply.

Command groups
~50
Releases
12 in 40 days
Platforms
macOS · Linux
Read the case study →
04

Reasoning Bank

Internal
The Reasoning Bank cycle: an agent session is traced, a judge turns the trace into a candidate lesson, a person accepts or rejects it, accepted lessons go into the bank, and agents recall them at the start of the next task. Agent sessionon Foundry Traceevery step kept Judgescores the attempt Candidatedo or avoid, and when A personaccepts or rejects The bankreviewed lessons recalled over MCP at task start and when a tool fails

A reviewed memory of what worked and what failed when agents operate Foundry.

Every agent session is traced and judged into a candidate lesson. A person keeps or drops each one. Agents recall the kept lessons when a task starts and when a tool fails.

Right lesson found
96%
Without reranker
68%
Kept by review
15 of 26
Read the case study →

Building something agents should run?

Raava Solutions takes on Palantir Foundry and AI agent work for clients. Tell us about the operation.

contact@raavasolutions.com