Agents write a lot of the code. I decide what gets built and I own what ships. Click through the loop I follow.
→→→→
I decide what to build. Before any code exists I write down the intent, the constraints and what done looks like.
Independent tech projects
Four systems I designed and built
Filter by technology to see which projects use it. Open a project, then click through its flow to see how the pieces connect.
Technology
Most proud project · Live product
Recentriq
A functional ERP for small businesses. CRM, projects, finance and HR sit in one workspace, so a lead becomes a project and then an invoice without anyone typing it twice.
How a deal moves through the product
→→→→
Inquiries enter the CRM with their source, a qualification status and CSV import for bulk lists.
What is inside
Communicate
Chat channels and threads, meetings with room allocation, bulk email on each company's own SMTP server.
CRM
Leads, clients, service orders and invoicing with taxes, discounts and recurring schedules.
Projects
Deliverables, task boards, knowledge bases, and digital docs with version control and share permissions.
Team
People records, departments and schedules, leave categories with paid and unpaid approvals.
Books and inventory
Chart of accounts, journals, expenses, stock and warehouses, purchasing with RFQs and multi-level approvals.
Settings
Module access, user permissions, billing, branding, notifications and security for each company.
Every company gets its own database schema, which keeps customer data separated at the database level.
Built to be bent to the business: modules switch off, fields rename, custom record types and approval routes can be added, and data imports from CSV.
I also run its go-to-market: a docs site, more than 90 articles and 10 head-to-head comparison pages against tools like Odoo, NetSuite and HubSpot.
Stack
Next.js, TypeScript, PostgreSQL, Drizzle ORM
Role
Founder. I designed, built and run it.
Status
Live. $5 per user per month with a 15-day free trial.
Business documents, cleanSyntheticNo errors on this set
2.31%
0.00%
Real-world wordsfigshare85.5% fewer errors
70.15%
10.17%
Handwritten linesBN-HTRd80.3% fewer errors
156.14%
30.75%
04080120160%
99.3%Words correct through the spot-then-read pipeline on 18 synthetic pages. Base model: 55.1%.
728 / 728Amounts and dates found on those pages.
0.15%Error on 12 full Bangla pages, with a box for every line. English, 6 pages: 0.12%.
650,211Training examples seen by the chosen checkpoint.
Handwriting is still the weak spot. At 30.75% the fine-tune is far better than the base model, and the model card advises against it for handwritten documents. All results come from public and synthetic sets, and the model has not been tested on a hand-labelled set of real business documents. The model card tells readers to verify extracted figures. Numbers are from the model card.
LoRA fine-tuning at rank 16 and alpha 32 on the language model's attention and MLP layers. The vision encoder stays untouched.
Training data mixes my own synthetic pages and lines in 18 Bangla and English fonts with public sets: BN-HTRd handwriting, IIIT Hyderabad Mozhi, Bangla Wikisource, figshare printed words and others. The chosen checkpoint saw 650,211 examples.
Evaluated by character error rate after Unicode normalization, on test sets held out from training.
Released as merged BF16 weights, a 24 MB PEFT adapter and GGUF files that run in llama.cpp.
Base
PaddleOCR-VL-1.6 (Apache 2.0)
License
CC BY-SA 4.0
Result
Printed-word error rate fell from 87.74% to 0.81%. See the results chart above.
A stock market data analytics platform that generates insights for investment analysts, so they can make effective investment decisions.
From market data to a decision
→→→→
Dhaka Stock Exchange market data is the raw input for everything the platform does.
Built for investment analysts who need answers drawn from market data, not another data export.
Python and FastAPI on the backend, TypeScript on the front end.
Stack
Python, FastAPI, TypeScript
Audience
Investment analysts
Status
Private. I can walk through the architecture and data model in an interview.
Analytics platform · Private
Hotel Analytics
A platform that reads customer reviews automatically, scores their sentiment, finds the pain points guests keep running into and helps hotels fix them.
From reviews to a better stay
→→→→
Guest reviews are the input. Nobody has to read them one by one.
Automated review sentiment analysis across all of a hotel's reviews.
Pain-point discovery, and a path from each pain point to a fix, to improve the overall customer experience.
Python on the backend, TypeScript on the front end.
Stack
Python, TypeScript
Audience
Hotels and hospitality teams
Status
Private. I can walk through the architecture and data model in an interview.
Job experience
Where I have worked
Retail · Bangladesh
ACI Logistics Ltd
Research Executive
ACI Logistics operates Shwapno, Bangladesh's largest retail chain, with more than 1,000 outlets across the country. Behind the stores sits an end-to-end supply chain that runs from growers to the shelf, selling fresh produce, groceries, household goods and more.
Role
Research Executive
Brand
Shwapno
Reach
1,000+ outlets across Bangladesh
Supply chain
End to end, from growers to the shelf
Technology services
Fame IT Networks Ltd
Business Analyst
An IT and technology services company with sister concerns in telephony and co-working. Client delivery, procurement and the systems behind them gave me a close view of how businesses actually use software.
Role
Business Analyst
Services
IT and technology services
Telephony
Fame PBX
Co-working
FameX 24/7 Workspace
Working across time zones
Where your day lands in Dhaka
Choose your time zone and your working hours. The tool shows what those hours are in Dhaka.