01 · About
Self-taught.
I'm a developer from Dhaka. I started writing code in 2018, self-taught, on a phone. Web first, deep on SvelteKit and Go, then in mid-2026 I moved into LLM pretraining and trained Alo-70M from scratch on TPU v5e-8.
I've been obsessed with AI since GPT-2 was the biggest open-source model around. The goal is simple: bilingual Bengali models that are open weights, self-funded, and small enough to run anywhere. The web work pays for it and ships it.
Currently pretraining a bilingual LLM in the 1B parameter range . Self-funded, fully open source.
Facts
- Location
- Dhaka, Bangladesh
- Focus
- Bengali NLP · web
- Coding since
- 2018
- Founded
- SurjoLabs (non-profit)
- Training
- Self-taught
Stack
Web
- · SvelteKit
- · TypeScript
- · Tailwind CSS
- · Go
- · PocketBase
- · Vite
ML
- · PyTorch
- · transformers
- · Hugging Face Hub
Infrastructure
- · Docker
- · Netlify
- · Vercel
- · Git
- · Linux
02 · Work
Selected work
02
SurjoLabs
Non-profit organization for bilingual Bengali models. Open weights, no funding, no paywalls.
03
PaymentTracker
In-production education management platform for a coaching center. Payments, students, and daily operations, beyond just tracking fees.
05
Pathway
Turns an Android phone into a local SMS and USSD gateway. No cloud, no server, no subscriptions. Pure device-side networking.
07
Omega-0
Design-focused web application built with attention to interaction and full-stack detail.
03 · ML model card
Alo-70M, evaluated
Zero-shot, continuation-based log-probability evaluation across seven Bengali reasoning and knowledge benchmarks. Every number below is published in the model card. The 69M-parameter model beats the 270M-parameter Gemma baseline on two of seven.
| Benchmark | Alo-70M(SFT) | Gemma-3-270M | TigerLLM-1B | Alo-70M-Base |
|---|---|---|---|---|
| bangla_mmlu_bn | 26.29% | 26.81% | 27.66% | 26.31% |
| bangla_commonsenseqa_bn | 25.88% ▲ | 22.77% | 25.14% | 28.42% |
| indicbench_arc_bn_challenge | 24.15% | 25.34% | 27.13% | 22.70% |
| boolqa_bn | 48.70% | 51.30% | 52.40% | 48.42% |
| openbookqa_bn | 30.58% | 31.99% | 34.21% | 31.39% |
| piqa_bn | 50.05% ▲ | 49.51% | 49.51% | 50.49% |
| hellaswag_bn | 26.89% | 27.85% | 31.01% | 27.27% |
Alo-70M vs Gemma-3-270M
accuracy %bangla_mmlu_bn
bangla_commonsenseqa_bn
indicbench_arc_bn_challenge
boolqa_bn
openbookqa_bn
piqa_bn
hellaswag_bn
Wins on bangla_commonsenseqa_bn and piqa_bn.
Size gap
3.9x fewer parameters than Gemma-3-270M
Wins
2 of 7 benchmarks, at 69M
Inference
CPU and phone, no GPU needed
Architecture
- Layers
- 12
- Hidden
- 512
- FFN
- 1408
- Attention
- GQA 8 / 4
- Positional
- RoPE
- Embeddings
- Untied
- Context
- 1024
- Parameters
- 69M
Training
- Hardware
- TPU v5e-8
- Optimizer
- Fused AdamW
- LR schedule
- Cosine, peak 3e-4
- Epochs
- 3
- Batch
- 32 (grad-acc 4)
- Precision
- AMP
- Format
- ChatML
- License
- Apache 2.0
Tokenizer
- Custom Bengali BPE, built for compact subword coverage.
- ChatML formatting via the tokenizer's chat template.
Alignment tax
SFT aligned the model for stable generation and instruction following, but it cost measurable zero-shot reasoning: CommonsenseQA dropped from 28.42% to 25.88%. Structured-formatting samples (12.5K) were deliberately excluded from the mix; they crushed a sub-100M model's capacity. At 69M parameters the model is built for text processing (editing, summarizing, extraction), not encyclopedic retrieval. It runs on CPUs and phones.
Currently pretraining a bilingual LLM in the 1B parameter range
Self-funded, fully open source
04 · Journey
Hobbyist to pretraining
2018
First lines of code
Started writing code on a phone, after school. Self-taught, building whatever came to mind.
2019 · 2020
The AI bug
Got hooked on LLMs when GPT-2 was the biggest open-source model around.
Sep 2024
First portfolio ships
Self-taught web development. Started with SvelteKit and never looked back.
2025 · Mid
Fullstack deep-dive
Go, PocketBase, TypeScript. Shipped experimental apps and utility tools for the local market.
Nov-Dec 2025
First utility products
attendly, OCR, starter templates. Small tools with real users.
Jan-Feb 2026
Pathway + Pollin Coder
Built a serverless Android SMS gateway and became maintainer of a free AI coding platform.
Jun 2026
Govt Photo Resizer
A client-side tool for Teletalk photo standards. Thousands of applications prepared locally in the browser.
Jun-Jul 2026
The LLM era
Pretrained Alo-70M-Base from scratch on TPU v5e-8, SFT-aligned it on a curated Bangla instruction mix. It beats Gemma-3-270M on 2 of 7 Bengali benchmarks.
Aug 2026
SurjoLabs + 1B scale
Founded SurjoLabs, a non-profit for bilingual Bengali models. Started pretraining a bilingual 1B-parameter LLM, self-funded and open. PaymentTracker went live in a coaching center.
Fahad delivered exceptional work on our projects. His attention to detail and technical expertise with SvelteKit made all the difference.
Mats Dale Soma
Founder · Redveil
05 · Contact
Let's build something
LLM collaboration, SvelteKit work, Bengali NLP research, or just to talk shop. All channels below are open.