Agents that do more
Most agents answer. Lanni executes. Built by PersistentAI on top of FireFlow orchestration for high-stakes finance, where mission-critical can't mean conventional.
2
FireFlow engine
A platform for building AI agents from ready-made blocks, like a construction kit, platform can be deployed on-prem. Architecturally guarantees that every request is processed.
Ready-made templates, plus the ability to share and reuse components, enable fast product iterations. Integrates with any service via MCP, CLI or HTTP/S.
Learn more about FireFlowReady-made templates, plus the ability to share and reuse components, enable fast product iterations. Integrates with any service via MCP, CLI or HTTP/S.
3
Flame Chorus
The chat engine: manages chats, participants, and messages. Chat engine supports features like AG-UI, streaming, markdowns and its own library of components directly integrated in FireFlow.
Currently moving from test version to production.
Learn more about Flame ChorusCurrently moving from test version to production.
4
MemoryTree
Memory layer built into the platform: full conversation history, versioned files, structured data, RAGs and a log of every decision. Gives fine-grained control of memory and context across users, agents, and chats.
This makes memory explainable, auditable, and reusable across workflows while allowing agents to interpret it correctly.
Learn more about MemoryTreeThis makes memory explainable, auditable, and reusable across workflows while allowing agents to interpret it correctly.
PersistentAI
THE STACK·WHAT YOU GET
Three products, One execution layer
Lanni is built on top of an orchestrator designed for
deployment in regulated, transaction-heavy financial
environments
(FireFlow
) with an AG-UI based chat engine
(FlameChorus
) & a custom persistent memory layer
(MemoryTree
).
deployment in regulated, transaction-heavy financial
environments
(FireFlow
) with an AG-UI based chat engine (FlameChorus
) & a custom persistent memory layer (MemoryTree
).#1The Orchestrator
FireFlow
A context-aware router reads user intent and invokes the right specialised agent — pricing, eligibility, documentation, execution — in sequence, under one governed flow with hard limits. One conversational entry point connects to dozens of capabilities, with no manual handoff logic.
▸ Multi-agent orchestration with delegated sub-tasks
▸ Available actions, order & permissions locked outside the model
▸ Workflows pause for minutes or days and survive any failure
#2The Persistent Memory
MemoryTree
Memory is a stack layer, not a sidecar. Four built-in layers give every agent durable, queryable, auditable state — so it never forgets how it solved a problem and every change can be rolled back.
▸ Conversation history across every surface
▸ Versioned files — git-style branching & diff on lakeFS
▸ Structured data via FFDB · DuckLake — SQL, JSON, full-text, geo, vectors & graph in one workspace file
#3The Chat Engine
FlameChorus
The Agent-Generated User Interface protocol lets agents propose context-aware components — charts, product cards, forms, confirmations — rendered in real time inside the chat. Every element is a deterministic wrapper around existing node output ports, not probabilistically generated code.
▸ Rich interactive UI from a single chat thread
▸ Deterministic wrappers — never hallucinated markup
Persistent AI owns its entire horizontal stack. This is what enables Lanni - or custom Lanni-derived agents - to excel where wrapped solutions fall apart.
VFS, VMs and Sandboxing
FireFlow runs workflows in isolated OpenSandbox environments with per-process secrets and policies. A versioned file system (VFS) keeps full change history; secrets are vault-encrypted at rest.
Visual Flow Builder
Author agents in a drag-and-drop editor with 146+ pre-built nodes. Flows are versioned, diffable .fflow packages — institutional assets, not one-off integrations.
Replay & Auditability
When a regulator questions an outcome, replay the full decision step by step. Agents run sandboxed with only the permissions they need — a misbehaving agent can't reach beyond its scope.
Deterministic by design
Intent
User goal enters as a
single request.
single request.
Route
Orchestrator decomposes &
selects agents.
selects agents.
Execute
DBOS persists every step
before proceeding.
before proceeding.
Render
Generating relevant visual
elements based on user's intent
elements based on user's intent
Settle & Audit
Result rendered; full trail
retained.
retained.
Lanni decomposes user intents into plans, resolves them using all capabilities available within its composably defined parameters and compiles an immutable, auditable execution graph which it runs to completion.
Launch LanniComing Soon
Connects to anything machine-readable
Lanni’s execution stack doesn’t run in a silo. The platform includes three native extension paths:
- An MCP surface with automatic node generation and capability discovery
- A Zero-code HTTP request nodes for any REST API
- A granular, custom TypeScript nodes
MCP
REST / HTTP
TypeScript
PostgreSQL
lakeFS
Docker
tRPC
x402
One spine. Audience-specific proof.
A context-aware router reads user intent and invokes the right specialised agent — pricing, eligibility, documentation, execution — in sequence, under one governed flow with hard limits. One conversational entry point connects to dozens of capabilities, with no manual handoff logic.
The casual user
▸ Why care, in a market full of AI apps?
Every other assistant gives an answer and leaves the doing to you. Persistent plans the group trip and sends the invites, finds the flights, pins the map.
Advice vs. a finished plan — open and replayable, so you can actually trust it with the booking.
The power user & builder
▸ Why this over Hermes or rolling my own?
It grows with you like the best open agents and runs on a durable, auditable engine — so it doesn't stall on long, multi-step jobs.
No lock-in, fully inspectable, extensible (MCP, Skills, flows), self-hostable. The agent that finishes.
The institution
▸ Why FireFlow over building it?
The path alone is 2–4 years and $5M+; 40% of agentic projects are projected cancelled by 2027.
FireFlow ships the full lifecycle in one open architecture you run in your own perimeter. You don't become a node — you own the rails.
Leadership
Co-Founder & CEO
Mike Sarvodaya
Former prop firm & hedge fund quant & risk manager. Institutional digital asset strategy leader since 2016 & AI solutions architect since 2022.
Co-Founder & CTO
Vladimir Maslyakov
Tech leader with 27+ years of experience across fintech, finance & web3 including co-founding and scaling Blum Crypto to 100m+ users.
Co-Founder (Non-Executive)
Max Rabinovitch
Business strategist integral to the launch and scale-out out the Chiliz ecosystem and a leader in the emergence of SportFi as a vertical.
Explore how Fireflow can help
Conduct a 4 week assessment of the institution's current infrastructure, regulatory posture, and integration scope followed by a proposal for a staged deployment.
