US Fintech Secures $2M+ Funding With a Generative AI Financial Modeling Platform
A US fintech wanted analysts to build credible financial models from public filings in minutes rather than days. We built a retrieval-augmented AI platform with rigorous evaluation and a SOC 2-ready cloud foundation, and the working product helped the company close a funding round of more than $2M.
What stood in the way
The founding team had a clear thesis: equity analysts at small funds spend most of their time pulling numbers out of 10-Ks and 10-Qs before any real analysis starts. Their early prototype, a thin wrapper around a general chat model, looked impressive in demos but produced confident errors on revenue lines and segment data. For a financial product, a single invented figure destroys trust, and prospective investors were asking pointed questions about accuracy.
They also needed an architecture that would survive enterprise security reviews later, without spending seed money on a full compliance program before product-market fit. The build also had to be demo-ready for investor meetings already on the calendar, so the timeline left little room for detours.
How we solved it
We rebuilt the product around grounded retrieval and measurable accuracy rather than model cleverness. Every number the platform shows is traced back to a specific passage or table in a source filing, and every change to prompts, models or retrieval is scored against a growing evaluation set before it reaches users. The cloud foundation was designed from day one to map cleanly onto SOC 2 controls.
RAG over structured filings
A pipeline that parses SEC filings into sections, tables and XBRL facts, embeds them with metadata, and retrieves by company, period and line item so answers are grounded in the right document.
Citations on every figure
Each extracted number carries a link to its source passage and page. Analysts can verify any cell in one click, and figures that cannot be grounded are flagged instead of guessed.
LLM evaluation harness
An automated test suite of several hundred analyst-verified questions scores accuracy, citation quality and refusal behavior on every release, letting the team compare models and prompts with evidence.
SOC 2-ready cloud foundation
Infrastructure as code on AWS with tenant isolation, encryption, audit logging, least-privilege access and CI/CD controls, documented so a future SOC 2 audit starts from evidence rather than rework.
How the Project Unfolded
Accuracy baseline and design
Built an evaluation set with the founders, measured the existing prototype against it, and designed the retrieval architecture and data model for filings.
Ingestion, retrieval, modeling UI
Delivered the filing ingestion pipeline, grounded extraction with source citations, and a spreadsheet-style modeling interface that analysts could edit, extend and export.
Evaluation and hardening
Iterated on prompts, chunking and model choice against the evaluation harness, then added production monitoring, per-tenant cost controls and tenant isolation.
Pilot users and investor demos
Onboarded pilot analysts, fixed issues surfaced in real use, and prepared a reliable demo environment and technical documentation for investor diligence.
Results that mattered
The rebuilt platform reached 96% figure-level accuracy on the evaluation set, and pilot analysts cut the time to a first-pass model by roughly 85%. Just as important, the team could now show investors how accuracy was measured and how it improved release over release. That evidence, alongside a working product in the hands of pilot users, helped the company close a funding round of more than $2M.
The evaluation harness has become the team's main release gate. They now switch models and tune prompts with confidence, and security questionnaires from larger prospects are answered from existing documentation rather than rushed engineering work. The platform is now expanding to earnings call transcripts and private company data.
Explore the Capabilities
More Success Stories
GamingGaming Startup Takes an AR Mobile Game to 200K Downloads in 30 Days
Retail & E-CommerceE-Commerce Platform Scales From 10K to 1M Monthly Users With Zero Downtime
Healthcare & PharmaClinic Network Automates Patient Intake and Claims Paperwork With Document AI
Have a similar challenge?
Book a free discovery call. We will map your goals and show you how we would approach it — no obligation.
Let's Build Something Extraordinary
Tell us about your project — whether it's an AI system, a game, an AR experience, or a complete platform. We'll schedule a free discovery call and show you exactly what's possible.