Projects

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

Chai Kahan: An Offline-First Compass to the Nearest Chai Dhaba

An installable web app that points a truck-art compass at the nearest chai dhaba in Lahore. Open it once with signal to download the city pack; after that GPS, the magnetometer and 454 places in IndexedDB do the rest, with no network at all.

Problem

Finding a dhaba needs a map, and maps need signal.

Solution

Download the city once. The compass works with nothing.

Next.js 16

TypeScript

PWA · Serwist

IndexedDB

Tailwind v4

OpenStreetMap + Overture

Tell me more

View Workflow Details

Workflow Features

Compass needle from fused magnetometer and GPS heading, spring-animated

454 dhabas and cafés in one Lahore pack from OSM, Overture and a curated list

Five views inside one precached document, so a cold start boots offline

Service worker with network-first manifest and runtime-cached packs

Pack versioning with atomic swap and an update toast

Onboarding grants location and motion, then downloads the pack with progress

Arrival state under 25 m: needle spins, phone vibrates, open in Google Maps

Vitest unit tests and Playwright end-to-end runs, including offline reload

Business Impact

454 places in the Lahore pack

0 network requests after first install

Nearest place found on device in under 1 ms

Boots cold in airplane mode

Workflow Architecture

OSM + Overture + curated CSV

Classify · dedupe · build pack

Lahore pack (JSON)

IndexedDB on device

GPS + magnetometer

Bearing − heading

Truck-art needle

Dossier Bot: On-Demand Lead Briefings in Slack

A Slack bot that turns a name or a LinkedIn URL into a PDF lead briefing. An agent with memory and four tools decides whether to fetch an existing dossier or build a new one, and a nightly run briefs the top five leads before anyone asks.

Problem

Every call starts with twenty minutes of tab-opening.

Solution

Mention a name in Slack. Get a briefing.

n8n

Slack API

LangChain Agent

GPT-4o

Supabase

html2pdf.app

Tell me more

View Workflow Details

Workflow Features

Answers an @mention in-thread

Conversation memory across turns

Cache-first: checks the bucket before generating

Lookup by profile URL or by name

On-demand generation as an agent tool

Briefing written from the lead's own recent posts

PDF rendered, stored and linked

Nightly pre-briefing of the top five leads

Business Impact

Briefing on demand, without leaving Slack

Top 5 leads pre-briefed nightly

Cached before generated, no duplicate work

One shared PDF instead of scattered notes

Workflow Architecture

Slack @mention trigger

Session parsing

Agent with memory buffer

Bucket lookup for existing dossier

Lead search by URL or name

On-demand generation tool

Agent writes from profile + posts

PDF render, store, reply in-thread

Zippit AI Content Engine: Automated SEO Blog System

Replaced a $1K/month content agency with a $50–100/month automated pipeline producing 2+ bilingual SEO articles daily for an 80,000+ customer Dutch health brand.

Problem

A thousand dollars a month for two articles a week.

Solution

Two SEO articles a day. Two languages.

LLM Content Chain

Perplexity

Pinecone

RAG

Shopify GraphQL API

Translations API

Tell me more

View Workflow Details

Workflow Features

2+ bilingual articles daily (EN + NL)

Automated topic sourcing

H1/H2 structure + internal links

Auto meta descriptions

EEAT citation layer

Pinecone RAG grounding

Native Shopify publishing

Single review checkpoint

Business Impact

10–20x cost reduction

$1K/mo agency → $50–100/mo API

15–30% MoM organic growth

30+ articles/month automated

20+ hrs/week reclaimed

Workflow Architecture

Topic + source ingestion

Structured LLM drafting chain

Perplexity citation enrichment

Pinecone RAG grounding

EEAT quality layer

Human review checkpoint

Shopify GraphQL publishing

Same-URL multi-language output

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

Equity Pulse: Automated Portfolio Intelligence for Retail Investors

Built an end-to-end portfolio analysis engine for retail investors, OCR ingestion, per-holding fundamental enrichment, Graham Fair Value scoring, and AI-written PDF reports delivered by email.

Problem

Your broker shows prices. Not answers.

Solution

Screenshot in. Advisor-grade PDF out.

GPT-4o Vision

GPT-4o-mini

Yahoo Finance API

Google Sheets

html2pdf.app

SendGrid

Tell me more

View Workflow Details

Workflow Features

Screenshot OCR ingestion

Batch portfolio processing

Per-holding enrichment

Graham fair-value scoring

Beta-based risk flags

Human-in-the-loop review mode

AI-written PDF reports

Email delivery

Business Impact

Hours → under 10 min per portfolio

Zero manual data pulls

Advisor-grade HITL review

Parallel multi-portfolio runs

Workflow Architecture

Portfolio screenshot upload

GPT-4o vision extraction

Yahoo Finance enrichment loop

Fair-value + risk scoring

Google Sheets HITL gate

GPT-4o-mini report writer

html2pdf conversion

SendGrid email delivery