Domain
Application Development
Technology
AI + Custom NLP Engine
Industry
Fintech
Project Duration
2023-Ongoing
For Aw Ming Sheng, the frustration was personal. As a retail investor, he was tired of AI tools built for Wall Street, where answers were vague, sources were missing, and real insight was buried in a 100-page SEC filing. If you weren’t an analyst, you were left guessing.
He didn’t want another chatbot. He wanted a platform that could parse 10,000+ filings and give plain-English answers backed by real sections—so everyday investors could finally understand what they were buying. That vision became FundFlicks.
Most existing financial tools were trained on open internet text—offering generic answers with zero traceability. Worse, the tools assumed users were fluent in financial jargon. Asking about something as simple as an expense ratio often meant decoding legalese or navigating endless PDFs.
If FundFlicks was going to work, it needed to solve three things at once: deliver fast answers, ground every response in a verified source, and speak in the language real investors actually use.
Amit, our developer, knew that users wouldn’t judge the platform on its infrastructure; they’d judge it on how it spoke to them. So, the conversation layer wasn’t treated like output. It was treated like an experience.
To make that happen, he fine-tuned GPT-4 prompts to reflect financial context, ensured answers used natural cadence and clarity and, most importantly, mapped user intent with empathy.
Asking “What’s the risk?” shouldn’t trigger a glossary. It should feel like the system understands what’s really being asked and replies like a trusted analyst, not a search engine.
Assuming ChatGPT could “do it all”
Use GPT-4 only where it made strategic sense. Build custom NLP and validation layers to handle what LLMs couldn’t do reliably.
Slow AI ≠ useful AI
Design a high-speed pipeline with Redis, Celery, and smart async routing to deliver answers in under 2.5 seconds..
Answers with no proof
Ensure every answer links directly to its source in the filing, so users trust what they see.
Just loading raw SEC filings into a chatbot
Build a structured, searchable knowledge base using Pinecone - indexing 10,000+ filings for real semantic retrieval.
Users ask natural-language questions — no keywords, no filters, no forms. For example: “What’s the expense ratio of this fund?” These questions are captured through a Next.js frontend and seamlessly routed to the backend via Django REST APIs.
The query is interpreted by a custom NLP engine powered by GPT-4 and further enhanced with Perplexity’s API for contextual grounding. It’s not just parsed — it’s truly understood in a financial context.
FundFlicks runs semantic vector search using Pinecone, scanning over 10,000 pre-embedded SEC filings to locate the most relevant text blocks.
Retrieved content passes through a summarization and validation pipeline to generate an investor-friendly response in plain English, backed by references from the actual filing.
Heavy tasks like API calls and document processing run in parallel via Celery workers and Redis queues, keeping the user experience fast and responsive even under load.
User queries were first parsed by a custom NLP layer that identified the question type—such as metrics, risk, or compliance—before passing it to the model for accurate handling.
Prompts used financial context to guide GPT-4 toward clear, relevant answers while avoiding guesswork, jargon, or unnecessary complexity.
The system used Perplexity to enrich vague queries. For example, “fees” triggered clarification between management, platform, or distribution without extra input.
Before being shown, each answer passed a validation layer that removed uncertainty, added source citations, and ensured every response was clear, direct, and human-like.
The Result
Average time to answer investor queries achieved through
Redis queues, Celery workers, and async pipeline design.
On source-backed responses validated through real queries
against SEC filings.
Indexed for semantic search making FundFlicks one of the
largest retail-accessible NLP platforms for SEC data.
Offered to the VE developer by the client turning the build into a
partnership, not just a project.
