← BACK TO PORTFOLIO

INVESTLY

Arabic participatory investment platform that connects investors with entrepreneurs and startups — with an AI service that turns long project pitches into short, trustworthy summaries.

RoleLead Developer / ML Engineer
Timeline2025 — 2026
StatusCompleted — Graduation Project
StackASP.NET Core, React Native, Python, FastAPI, AraT5

OVERVIEW

Investly is my graduation project, completed and defended in 2026. It is an Arabic platform for participatory investment: entrepreneurs and startups publish their projects, and investors browse, compare, and fund them. The wider goal is to support digital transformation, entrepreneurial growth, and financial literacy in the region.

I led development and built the machine-learning side of the platform. Its centrepiece is an Arabic summarization service that condenses a long project description into a short summary — without ever inventing the numbers an investor relies on.

THE PROBLEM

  • Long, uneven pitches. Every entrepreneur describes their project differently and at length, so comparing several projects quickly is hard.
  • Arabic text first. Descriptions are written in Arabic, which most off-the-shelf summarization tools handle poorly.
  • Numbers must stay exact. A summary that rounds, swaps, or invents a funding figure is worse than no summary at all.
  • AI must not take the platform down. If the model is slow or unavailable, browsing and investing have to keep working.

WHAT A PROJECT CARRIES

Each listed project combines a free-text description with structured fields stored in the database. That split is what the summarization design is built around.

Funding
Funding goal, amount raised so far, and the percentage funded.
Terms
Campaign duration, expected return, and the ownership share offered to investors.
Context
City and sector, so investors can filter and compare like with like.
Description
The entrepreneur's own Arabic write-up of the idea, the market, and the plan — the part that gets summarized.

HOW A SUMMARY IS PRODUCED

01RequestThe investor asks for a summary from the project page. The ASP.NET Core backend validates the request and sends the description to the summarization service over HTTP.
02Clean the Arabic textThe service checks the text is present and suitable, normalizes whitespace, and removes diacritics where needed.
03Split and scoreThe text is split into sentences and words, and the most relevant sentences are identified.
04SummarizeThe extractive engine builds a summary from the selected sentences — or the AraT5 model writes one when its trained weights are available.
05Attach verified factsThe backend appends the project's key figures straight from the database: goal, raised amount, funding percentage, duration, expected return, ownership share, city, and sector.
06DisplayThe app shows the summary alongside the original description, which always stays one tap away. If the service fails or times out, a temporary message appears and the rest of the platform carries on.

FOUR DESIGN DECISIONS

An isolated AI service
Summarization runs as a separate Python / FastAPI service behind a request timeout. It can be stopped, upgraded, or replaced without touching the core platform.
Two engines, one fallback
Extractive summarization is the default because it needs no trained model and cannot hallucinate. AraT5 adds abstractive summaries when its weights load; if it fails or returns unusable output, the service falls back automatically.
Numbers are never generated
Financial facts come from the database, not from model output. Keeping them out of the generated text removes the risk of a summary changing or inventing a figure.
It informs, it does not decide
The summary shortens reading time only. It never recommends accepting or rejecting a project, and investors are pointed back to the full description before deciding.

OTHER AI FEATURES

  • Face verification: confirms an investor's identity during onboarding.
  • Document forgery detection: analyzes uploaded legal documents for signs of tampering.
  • Keyword extraction: pulls the key terms out of long startup pitches.
  • Adaptive UX: personalizes the dashboard based on how each user behaves.
  • Two interfaces: a React Native app for end users, plus a Tkinter desktop tool for working with the models directly.

HOW QUALITY IS JUDGED

The evaluation plan scores the summaries on a test set kept separate from development data, combining automatic metrics with human review:

ROUGE-1 / 2 / L
Overlap with reference summaries. Useful, but a high score alone does not prove the meaning is right.
Precision, Recall, F1
How much of what the summary states is correct, and how much of what matters it actually covers.
Field Score
The share of the project's real fields that appear correctly in the summary.
Human review & latency
Clarity, whether key information survived, whether anything was added — plus average response time.

WHAT I TOOK AWAY

The model was the easy part to get excited about; the design around it mattered more. Separating facts from generated text, giving the AI a safe fallback, and keeping it out of the platform's critical path are what make a feature like this trustworthy enough to put in front of people making financial decisions.

Note: Investly was completed and defended as my graduation project in 2026. The repository is private; source code is available on request.