Six production AI systems — satellite flood intelligence, contract risk auditing, admissions equity, agricultural advisory, cricket analytics, industrial inspection. All built on free-tier infrastructure. All designed around the actual constraints of South Asia, not Silicon Valley.
The standard playbook for "AI for development" is to take a mature Western system, translate it, and deploy it at scale. It fails because the infrastructure assumptions don't transfer. Bangladeshi garment factories don't have $12,000 optical inspection rigs. 78% of the rural population owns a button phone, not a smartphone. Government documents exist as degraded photocopies of mixed Bengali-English text.
My approach is different. I study the best-in-class solution in each domain, identify its infrastructure assumptions, and re-architect from first principles for the actual constraints of South Asia. That process consistently produces systems that are 10–80× cheaper to deploy while maintaining competitive accuracy — because the constraints force genuine innovation, not translation.
AI is the most powerful tool in human history that remains almost entirely unavailable to the people who need it most. That is a design problem. I build the designs.
"The insight from the garment factory: India sells ready-made optical inspection for crores. I rebuilt it on a recycled Android phone at $150 CAPEX. The factory doesn't care who made the component — it cares about the margin."
I'm Samiul Karim — a self-taught applied ML engineer based in Dhaka, building production AI systems for domains that established labs ignore because the TAM looks small from San Francisco. From Dhaka it looks like 170 million people.
My background is unconventional: deep domain knowledge in garments manufacturing, Bangladeshi agriculture, public health surveillance, cricket analytics, and government document processing — combined with a Princeton University network built during an extended period as a Princeton alumni interviewer.
I work in long focused sessions. Every system in this portfolio was built to production-grade standards — proper authentication, eval harnesses, retry logic, secrets management — because demo-grade code teaches the wrong habits.
Reach me at sam@stitchmark.space
Each project below was built to production standards. None required paid cloud credits to reach a working prototype.
All projects are technically proven on free-tier infrastructure. The engineering problems are solved. What scales impact is serving capacity — RAG corpus retrieval at volume, Gemini API at production query loads, Cloud Run autoscaling.
Projects are ranked below by social impact. Any credit allocation advances the work. Larger allocations unlock proportionally greater reach.
AI agricultural advisory for 57 million rural Bangladeshis with no smartphone access — delivered over SMS and USSD at $0.001 per query. The constraint is not the AI. It is the last mile.
Greater allocation adds: full Bengali NLP pipeline, multi-crop seasonal corpus, carrier USSD deployment.
Near-real-time pixel-level flood maps from Sentinel-1 SAR for government disaster response. All-weather. No human annotation. No manual steps between satellite overpass and map delivery.
Greater allocation adds: 4-month operational pilot, active learning loop with SPARRSO validation, public dashboard.
Forensic financial aid intelligence for high-need Bangladeshi students navigating elite US admissions — replacing generic chatbot advice with verified CDS audit data and structured visual counsel.
Greater allocation adds: full corpus (15+ universities), Bangladesh-specific curriculum contextualization overlays.
Contract risk intelligence with adversarial self-auditing. Bringing the evidentiary standards of a senior partner to every junior associate in Dhaka. Already deployed. Needs production serving scale.
Greater allocation adds: Bangladesh-specific corpus, Bengali-language query handling, SaaS tier launch at $149/month.