Equity Pulse: Automated Portfolio Intelligence for Retail Investors

The Problem

Most retail brokerage platforms give investors access to global markets but zero analytical tooling. Getting a meaningful view of a 15–20 stock portfolio, current valuation vs fair value, analyst consensus, upcoming earnings, risk profile, means manually pulling data across Yahoo Finance, news aggregators, and brokerage research tabs. For an independent advisor running the same review across multiple client portfolios, that's hours of work per cycle that produces nothing proprietary.

What I Built

An automation pipeline that takes a brokerage portfolio screenshot and returns a PDF investment report, no manual data entry required. OCR Ingestion, GPT-4o vision extracts every holding (stocks and ETFs) directly from a brokerage portfolio screenshot or export, including ISINs and position sizes. A batch ID groups each upload so multiple portfolios run in parallel without collision. Per-Holding Enrichment, Each ticker runs through a Yahoo Finance enrichment loop: live pricing, analyst consensus ratings, recent news, earnings calendar, and beta coefficient, all pulled and structured automatically. Graham Fair Value Scoring, Intrinsic value calculated per stock using √(22.5 × EPS × Book Value) × 1.15, surfacing whether each position is trading above or below fundamental fair value at a glance. Beta-Based Risk Flagging, Every holding is tagged Low (β < 1.0), Normal (1.0–1.2), or High Risk (β > 1.2) to give the portfolio a quick risk-distribution read. Two Delivery Modes:

Direct, fully automated end-to-end. Portfolio upload → enrich → report in under 10 minutes.
HITL (Human-in-the-Loop), enriched data lands in a Google Sheets review interface first. The advisor confirms before the PDF generates, allowing annotations or corrections before client delivery.

AI Report Generation, GPT-4o-mini writes a structured HTML investment report across all holdings, converted to PDF via html2pdf.app and delivered as an email attachment via SendGrid.

Outcomes

Per-portfolio analysis time: hours → under 10 minutes
Full fundamental snapshot (valuation, analyst view, news, risk) on every holding with zero manual data pulls
Advisor-grade review workflow via HITL mode before client-facing delivery

Also view other projects

Excelr8: AI-Powered GTM Automation Platform

Built the full GTM automation dashboard covering lead import, enrichment, campaign setup, CRM sync, and KPI reporting, cutting campaign launch time from hours to minutes.

Problem

Leads live in spreadsheets. Campaigns launch by hand.

Solution

Campaign setup, hours down to minutes.

n8n

HubSpot

Apollo

Clay

LinkedIn Signals

LLM Pipelines

Tell me more

View Workflow Details

Workflow Features

Multi-stage lead enrichment

Qualification scoring

LinkedIn Post Radar

Warm-lead auto-sequencing

Bi-directional HubSpot sync

Dedup + error recovery

AI newsletter pipeline

Real-time KPI dashboard

Business Impact

Campaign launch: hours → minutes

14–30 hrs/month CRM entry eliminated

10–36 hrs/month content automated

500–2,000 leads/month reported live

Zero manual reporting

Workflow Architecture

Apollo / Clay lead import

Enrichment + scoring stages

LinkedIn signal monitor

Sequence trigger engine

n8n HubSpot sync layer

LLM newsletter chain

KPI aggregation + dashboard

Arabic OCR Post-Correction: A 0.5B Model That Makes Scanned Archives Searchable

A 0.5B language model, fine-tuned with LoRA on 57k synthetic pairs, that repairs Arabic OCR output. Word error rate falls from 41% to 20% on held-out text. Trained in 80 minutes on a laptop GPU, and it runs on CPU, on-premise.

Problem

Scanned Arabic archives are indexed and still unsearchable.

Solution

A 0.5B model repairs the OCR. Word errors halved.

Qwen2.5-0.5B

LoRA / PEFT

PyTorch

Hugging Face

Python

Tell me more

View project

View Workflow Details

Workflow Features

Confusion model weighted by how Arabic script actually fails

Dot and skeleton confusion, word splits and merges, dropped and spurious characters

Severity sampled 4–18% per example, so one model covers clean and degraded pages

LoRA r=32 on all attention and MLP projections, bf16, no quantisation

Greedy decoding: one right answer, no sampling

Guardrail rejects wholesale rewrites and keeps the original

CER / WER evaluation by severity band against the raw-OCR baseline

Trains on a free Colab T4 with an automatic fp16 fallback

Business Impact

Word error 41.1% → 20.4%

Character error 8.1% → 6.7%

80 minutes on one laptop GPU

Runs on CPU, on-premise

Workflow Architecture

Clean Arabic news corpus

Script-aware corruption model

57k noisy → clean pairs

LoRA finetune of Qwen2.5-0.5B

Greedy decode

Drift guardrail

CER / WER scoring