FundersAI

• 

  FundersAI - Making Indian Stock Market and Mutual Fund Research Easy

Inspiration

Indian stock and mutual-fund research is often scattered across dashboards, spreadsheets, fund factsheets, and unreliable summaries. I wanted to build a research workspace that makes analysis easier without pretending to be a financial advisor.

FundersAI was inspired by a simple principle: AI should help people understand financial information, while deterministic calculations and official sources remain responsible for the facts.

What it does

FundersAI is a research-first web app for Indian stocks and mutual funds.

Users can:

• Compare mutual funds using returns, NAV history, alpha, beta, Sharpe ratio, drawdown, holdings, risk labels, and other structured metrics.
• Ask research questions through an AI chat interface.
• Explore results through an interactive comparison canvas.
• Read cited evidence from official AMC documents.
• See data freshness, coverage limitations, and source metadata.
• Receive an explicit abstention when the available evidence is insufficient.

FundersAI is designed for research and education. It does not execute trades or provide personalized investment advice.

How we built it

I built FundersAI with a Next.js and TypeScript frontend, a Python/FastAPI backend, Supabase/PostgreSQL for structured data, Cloudflare R2 for raw documents and archives, and GitHub Actions for scheduled ingestion and maintenance workflows.

The application uses a Supabase-first runtime so query-critical data can be served from normalized snapshots and caches instead of making uncontrolled provider calls during every request.

For mutual-fund research, official AMC documents are acquired, stored, parsed, validated, and indexed. The evidence layer supports deterministic lexical retrieval, optional semantic retrieval with embeddings and pgvector, cited excerpts, claim-level support checks, and abstention when evidence is missing.

The comparison system keeps calculations deterministic. AI is used for routing, explanation, and structured extraction, but it is not allowed to invent metrics or silently fill data gaps.

Challenges we ran into

The hardest part was not building a chatbot. It was building reliable financial data pipelines around inconsistent real-world sources.

AMC documents change layouts, contain PDFs with difficult tables, and sometimes expose incomplete or temporary links. A document can download successfully and still parse incorrectly. We had to introduce explicit ingestion states such as parsed, parsed_partial, needs_review, and failed so uncertainty was visible instead of hidden.

Another challenge was preventing AI-generated answers from sounding more confident than the evidence. We added source metadata, freshness information, coverage explanations, deterministic calculations, citation validation, and explicit abstention behavior.

We also had to balance retrieval quality against latency and provider cost. The system keeps a provider-free lexical fallback and treats semantic retrieval and reranking as measured, opt-in capabilities rather than assumptions of quality.

Accomplishments that we're proud of

The accomplishment I am most proud of is the complete path from official document to useful, cited research:

official AMC document → stored and parsed evidence → indexed chunks → retrieved sources → supported answer or abstention

FundersAI also has a working authenticated workspace with chat persistence, comparison workflows, scheduled data pipelines, an operations dashboard, and source-aware research surfaces.

The retrieval evaluation pipeline improved its development seed from 12/14 to 14/14 passing cases with the deterministic reranker. This is not presented as production proof; it gave us a repeatable way to measure changes before enabling more expensive retrieval components.

We also verified production official-document evidence queries that returned grounded answers with citations for investment objectives, benchmarks, risk levels, and expense-ratio sections.

Most importantly, the product makes limitations visible. A missing benchmark, incomplete document, stale result, or unsupported question is shown as a limitation instead of being disguised as certainty.

What we learned

We learned that trustworthy financial AI depends more on boundaries than on making the model sound intelligent.

Deterministic metrics should remain deterministic. Official documents need provenance. Retrieval needs evaluation. Partial data needs to be disclosed. And when evidence is unavailable, abstaining is often more useful than producing a plausible answer.

We also learned that reliability is an end-to-end problem. A parser, database schema, scheduled workflow, API contract, and frontend explanation must all agree about what a result means.

What's next for FundersAI

Next, we plan to:

• Expand reviewer-verified research questions beyond the initial development seed.
• Measure retrieval quality, latency, and cost on a larger official-document evaluation set.
• Complete hosted semantic-retrieval verification.
• Improve AMC document coverage and structured-field completeness.
• Reduce provider-backed explanation latency.
• Add more explainable research tools while preserving the research-only boundary.
• Continue improving the admin workflows used to detect, review, and repair unreliable source data.

The long-term goal is to make FundersAI a transparent research companion: useful enough to accelerate analysis, but disciplined enough to show its sources, calculations, and limits.

Built With

• agenticai
• cloudflarer2
• codex
• fastapi
• githubactions
• next.js
• python
• rag
• render
• supabase
• typescript
• vercel