01
Underwrites deals with you
A structured, scored assessment: observations, red flags, missing data, opportunities, a metrics check. It reads your model (native or uploaded Excel) as the source of truth and reasons from that, not a guess.
LucidAI · built into LucidOS · ⌘J
The analyst who already
knows your portfolio.
A conversational partner wired directly into your deals, properties, people, and capital. Ask it anything. Get a real answer, grounded in real numbers. Not a chatbot bolted onto a dashboard. An operator that lives where you work.
Connected to your operations
Yes. It reads your live AppFolio data.
LucidOS syncs eight AppFolio reports nightly, with a near-real-time refresh every 15 minutes for availability and rent roll. LucidAI reads straight from that synced data, so the answer reflects what’s actually happening at the asset, not a stale export someone pasted into a deck last quarter.
Every operating figure carries a freshness stamp
How's delinquency at Oak Ridge?
Total outstanding, the top delinquent tenants, and how fresh the number is.
Any open work orders on this asset?
The maintenance queue, by status — resolved, in progress, waiting.
What's the leasing pipeline looking like?
Showings, applications, leases signed, move-ins and move-outs.
What's rolling over this quarter?
Upcoming lease expirations across the property, on the horizon you ask for.
How's occupancy trending vs in-place rent?
The operating picture — occupancy against rent — in one breath.
What's the latest on this deal's property?
It resolves the deal → asset link for you and pulls the operations. No navigating.
Honest about latency. LucidAI tells you when the data was last synced. If a property hasn’t been connected yet, it says so instead of inventing a confident zero.
What LucidAI does
One assistant. The whole
data stack in reach.
02
Answers for the whole portfolio
Delinquency, work orders, leasing pipeline, expirations, occupancy, financials. The operating reality of every asset, on demand, from any screen.
03
Remembers everyone in the deal
Brokers, sellers, lenders, investors, tenants, prospects, and the full history of every interaction. “Who's the seller on Evaline, and when did we last talk?” It walks the relationships and answers.
04
Tracks the raise
Target vs committed vs funded, the investor pipeline by stage, who's soft-committed, what closed this week. All of it pulled from your capital-raise data, not reconstructed from memory.
05
Drafts the outreach
Campaign performance, workflow status, why a send failed. When you're ready, it drafts the email or automation for you to review. It proposes; you approve.
06
Acts, not just answers
Log an interaction. Create a task. Add a contact. Build a segment. Draft a deal. It turns a sentence into an action. It always proposes, then confirms, and only ever offers what you could do yourself.
Why it’s different
The difference between sounding right
and being right.
It’s grounded
It doesn’t make numbers up.
Most assistants will happily hallucinate a delinquency figure that sounds right. LucidAI won’t. It runs under a strict ground-truth precedence: computed figures and live data always win, and it is expressly forbidden from citing a stray memory or a fuzzy search hit as the source of a financial number. When it gives you a figure, it came from the system of record. When it doesn’t have one, it tells you, and offers to go get it.
It learns your firm
LucidAI builds a private profile of how your shop operates: your thesis, target markets, underwriting standards, terminology, key people. Over time it stops answering like generic software. Your corrections stick; your language becomes its language.
A memory that's actually useful
A real retrieval system over your org's knowledge (interactions, contacts, deals, call summaries, weekly rollups) with semantic search, so “what did we decide about that broker” finds the thread even when you don't remember the words.
Aware of where you are
It reads the page you're on. On a deal, it already has the model, returns, and people loaded. On a property, the operating snapshot. You never re-explain context. It's already in the room with you.
Under the hood
The engineering behind the answers.
LucidAI feels effortless because a lot of hard architecture sits underneath it. For the technically minded, here’s what makes the answers good.
Retrieval
Hybrid RAG, not a single fuzzy lookup
Every question triggers retrieval over your org's own knowledge, fused from two searches into one relevance score:
Semantic (vector) search
understands meaning — finds the right thread even when none of its words appear in your question.
Lexical (keyword) search
guarantees exact matches — a property name, a dollar figure, an email — never lost to paraphrase.
Embedded into 1024-dimensional vectors and indexed with HNSW approximate-nearest-neighbor, so recall stays fast from one property to a hundred.
Memory
Two layers, kept honest
Knowledge
facts drawn continuously from live data — re-embedding what changed, skipping what didn't (content-hash dedup), pruning what's gone, so the index never drifts from reality.
Long-term memory + firm profile
durable, distilled facts about how your org operates, extracted from usage and corrections, weighted by confidence, expired when stale.
This is what lets LucidAI answer in your firm's voice, with your thesis and terminology, instead of like generic software.
Grounding
Why it doesn't hallucinate figures
Retrieved text is background, not truth. LucidAI separates the two by design:
Structured context loaders
read the live record for the page you're on — the deal's model, the property's snapshot — from the system of record.
Function-calling tools (25+)
pull authoritative real-time data on demand: a deal analysis, a delinquency, an investor pipeline, a campaign's performance.
A strict ground-truth precedence governs it all: computed figures and live tool results outrank retrieved text or memory. Numbers come from the ledger; recall comes from retrieval; they never trade places.
Runtime
Governed tools, frontier models
Every action is a typed, validated tool with a propose-then-confirm contract. The write runs through the same permission-checked, audited path your UI uses. Nothing mutates without a human approval.
Anthropic Claude, streaming
answers arrive token-by-token, not after a long pause.
Cached prompt segments
your firm profile and platform instructions aren't re-billed every turn; heavy reasoning escalates to a stronger model only where it's worth it.
Feedback shapes memory
A team correction is weighted heavily and folded back into what LucidAI knows.
Freshness is tracked end-to-end
Operating answers carry a sync timestamp; never-synced assets are flagged, not faked.
Privacy is enforced at the index
A note marked confidential carries that restriction into the search layer — it can't resurface through a side door.
Technical snapshot
- Reasoning
- Anthropic Claude, streaming, with prompt-caching and model escalation for heavy tasks
- Retrieval
- Hybrid RAG — semantic vector search fused with lexical keyword search
- Embeddings
- 1024-dimensional, best-in-class model, incrementally re-embedded
- Vector index
- HNSW approximate-nearest-neighbor for fast recall at scale
- Memory
- Live knowledge chunks + distilled long-term memory + inferred firm profile
- Live data
- Function-calling tools + structured context loaders over your system of record
- Data source
- Your platform data, including AppFolio synced nightly (availability ~every 15 min)
- Actions
- 25+ typed tools, propose-then-confirm, permission-gated and audited
- Grounding
- Ground-truth precedence — figures from the ledger, never from a guess
- Security
- Per-org data isolation, role/permission inheritance, index-level privacy
How to think about the value
The same work, in one sentence.
Before LucidAI
With LucidAI
Open AppFolio, run a report, export, read
“How's delinquency at Oak Ridge?” — answered in the flow of work
An analyst spends an afternoon assembling a deal summary
A structured, scored analysis in seconds, ready to refine
“What did we decide with that broker?” → dig through email
Answered from the interaction history, with the person linked
Numbers pasted between tools, going stale on contact
One assistant reading the live system of record, timestamped
Generic AI that guesses
An assistant that knows your thesis, your assets, and your rules
Built for teams that can’t afford to be wrong.
Permission-aware
LucidAI never surfaces data a user can't see, and never offers an action a user couldn't perform themselves. It inherits your roles and permissions exactly.
Multi-tenant secure
Every answer is isolated to your organization at the data layer. A thread one teammate marked confidential never leaks into a shared answer.
Transparent about freshness
Live where it can be, honest about latency where it can't. You always know how current the number is.
FAQ
The questions operators ask.
Does LucidAI connect to AppFolio?
Yes. LucidOS syncs your AppFolio reports — rent roll, unit directory, delinquency, work orders, income statement, leasing summary, lease expirations — on a nightly cadence, with availability refreshed roughly every 15 minutes. LucidAI reads that synced operating data directly, so it can answer questions about delinquency, maintenance, leasing, occupancy, and financials for any connected property.
How current are the numbers?
As current as your last sync — and LucidAI tells you when that was on every operating answer. Availability and rent-roll data refresh near-continuously; the full operating picture refreshes nightly.
Will it make up a number if it doesn't have the data?
No. LucidAI is designed to refuse that. It sources figures from your system of record and, when it doesn't have the data, says so rather than guessing.
How does the RAG / search actually work?
LucidAI uses hybrid retrieval-augmented generation: your org's knowledge is embedded into vectors for meaning-based semantic search, and indexed for exact keyword matching, with the two fused into one relevance score. Retrieval runs on an HNSW vector index so it stays fast as your data grows. Critically, retrieved text is treated as background context — never as the source of a live figure. Numbers come from real-time tool calls and structured reads of your system of record.
Can it take actions, or only answer?
Both — but writes are always propose-then-confirm, and it will only ever offer an action you have permission to perform. Nothing changes without your approval.
Is my data private?
Yes. Answers are isolated to your organization, LucidAI respects your roles and permissions, and confidential notes never bleed into shared responses.
Does it work for acquisitions, not just management?
Both are first-class. Underwriting, scenarios, waterfall, comps, and deal risk on the acquisitions side; delinquency, work orders, and leasing on the operations side — and it connects the two, so a deal knows its asset.
Ask your portfolio anything
LucidAI doesn’t replace your team. It gives every person on it the memory of a ten-year veteran and the reach of the whole data stack, in one sentence.