Regulated investment platform
Pitch decks become drafted deals, and issuers approve each field before publishing
A live, regulated marketplace for private investment deals wanted AI to read issuers' decks, memos and financials. We built agents beside its .NET core that suggest each deal field next to its current value, and the issuer applies, edits or rejects it before anything publishes.
Financial ServicesPrivate investment marketplaceAI added to a live product
The challenge
Issuers copied deal data by hand from decks, memos and financials into long forms, inside a regulated product whose application and database could not change.
What we did
We built agents as APIs beside the .NET product that read the deal files, suggest each field for the issuer to apply, edit or reject, and draft the market summary and documents.
The outcome
The client approved the upload and extraction demo and the chat approach, confirmed several delivered agents working in its own testing, and took the APIs into acceptance testing.
- Inside
- A live, regulated private investment marketplace
- Reads
- Up to ten deal files per upload, in the background
- Each field
- Current value, AI suggestion: apply, edit or reject
- Delivered as
- Backend APIs for the client's own .NET team
The starting point
A regulated marketplace where issuers raise equity, debt or hybrid rounds, each tied to its own SPV, and investors browse deals, express interest, sign and track holdings. It already ran as a full .NET application on SQL Server with issuer, investor, broker and admin roles.
What had to change
Issuers copied deal data by hand from pitch decks, memos, financials and short pitch videos into long deal forms, page after page. The client wanted AI to fill the deal from the uploads while issuers stayed in charge of every value.
What could not break
- The .NET application and SQL Server database stay the regulated core
- AI reaches the product only through APIs, never the database directly
- Every suggested value waits for the issuer to apply, edit or reject it
- Chat has to work inside the existing .NET MVC screens, which had no native streaming
- Compliance and legal interpretation stay with the client
The calls we made
We kept the marketplace as the system of record and made the issuer the editor: the AI drafts, and a person decides, one field at a time.
- 01
Services around the core, not inside it
The agents run as FastAPI services in a separate AWS environment and reach the marketplace only through APIs, so the regulated application and its SQL Server database stay as they were.
Ruled outGiving the AI direct access to the regulated database.
- 02
Suggest field by field
On each page a popup shows the current value beside the AI's suggestion, and the issuer applies, rejects or edits it. Once a person finalizes a value, theirs is the source of truth.
- 03
The issuer's documents come first
The agents draw on the uploaded deal documents first, use web search only for market context and use the language model to write it up, so a deal reflects what the issuer provided.
- 04
A gate before publishing
A deal publishes only when the documents are uploaded or the issuer confirms going without them, the AI and human versions have been compared and a person ticks the confirmation.
- 05
One baseline agent, then the rest
At the client's request we finished the deal creation agent, with its chat and full workflow, first, as the standard every later agent follows, each with one narrow job and its own API.
How the system fits together
What we built
Deal creation agent
Reads up to ten pitch decks, memos, business plans and financials per upload in the background, shows the processing status and suggests values for the deal fields on each page, asking the issuer when minimum information is missing.
Field by field review
A popup puts the current value beside the suggested one, and the issuer applies, rejects or edits it. The AI version is stored next to the human one for the comparison before publishing.
Issuer chat
A context-aware chat that receives the current page, its fields and the question, then names the missing fields and drafts content such as use of proceeds. New screen fields do not need retraining.
Market summary page
An executive summary, competitor analysis, market dynamics and a performance chart with projections, written from the issuer's documents first and web search only for context.
Compliance and offering documents
Compliance and documentation template agents and a private placement memorandum generator that draft from the finalized deal, with templates kept under admin control.
How it works
The issuer uploads
Up to ten decks, memos, business plans and financials are attached to a new deal and processed in the background.
The agent suggests
Each page of the deal form shows the current value beside a suggested one, drawn from the uploaded files first.
The issuer decides
They apply, reject or edit each suggestion, and the chat points out fields that are still missing.
A person confirms
The pre-publish gate needs the documents or an explicit go-ahead without them, a reviewed comparison and a ticked confirmation.
Documents follow
The market summary and compliance and offering drafts come from the finalized deal, under templates the admin controls.
A closer look
{
"deal": "draft-3104",
"page": "Deal terms",
"files": ["pitch_deck.pptx", "financials.xlsx"],
"suggestions": [
{"field": "round", "current": "", "suggested": "Series A", "action": "pending"},
{"field": "stage", "current": "Prototype", "suggested": "Pilot production", "action": "pending"},
{"field": "pre_money_valuation", "suggested": null, "action": "ask_issuer"}
],
"publish_status": "awaiting_human_confirmation"
}What changed
The client approved the document upload and extraction demo and the chat approach, then confirmed several delivered agents working in its own testing. The regulated core and its database were never rebuilt, and only values a person accepts can reach a deal.
Under the hood
- Datics role
- AI added to a live product
- AI boundary
- AI reads only the documents uploaded for a deal and a few purpose-limited sources, suggests values, and never publishes or approves a deal or a KYC check.
- System of record
- The .NET application and SQL Server database stay the core, and finalized fields reach the database through the client's side, never directly.
- Models and data
- FastAPI services orchestrated with LangChain and LangGraph on Azure OpenAI, with SerpAPI web search for market context only.
- Deployment
- AI services on EC2 in a separate AWS environment beside the live application, designed so the client can move them into its own AWS account.
AI and models
Cloud and infrastructure
Apps and APIs
Also in the stack
How it shipped
Discovery and scope of work
Discovery calls, a proposal and a scope of work fixed the boundaries: APIs only, a separate cloud environment, and people approving every regulated step.
Walkthrough of the live platform
The client walked us through the live platform's roles, deal types, legal templates and data room, and named the pain: issuers copying deal data from decks by hand.
First issuer agent demo
Decks and other deal files were parsed into a structured deal template that filled the front end, with the issuer asked for anything missing.
Requirement alignment
Joint sessions settled AI deal creation, the side by side versions and the pre-publish gate, and made the deal creation agent with chat the baseline for the rest.
Handover and acceptance testing
The client approved the upload and extraction demo and the chat approach, and the agents went to its team as APIs for integration and acceptance testing.
APIs for their team
We delivered each agent as a backend API, and the client's own .NET developers owned the screens and the integration.
One integration branch
The client's developer had access to our AI repository, and integration ran through the client's own integration branch.
Demo, then hand over
Each agent was demoed to the client, and joint sessions settled the details before it went into acceptance testing.
What this unlocked
The client asked for ongoing maintenance and support, estimates for further development and a move of the AI services into its own AWS account.
Related work
Have a product with the same kind of constraints?
Tell us what already exists and what you want it to do. We will tell you what it takes.