Notes from building products, working with AI, and following interesting problems a little further.
Build your own anime-inspired portfolio with a copyable prompt for Codex, Claude Code or Cursor. My process for character identity, connected scroll scenes, transformations, sound, an AI assistant and mobile review.
Read the note ↗Build your own anime-inspired portfolio with a copyable prompt for Codex, Claude Code or Cursor. My process for character identity, connected scroll scenes, transformations, sound, an AI assistant and mobile review.
↗Launch a WireGuard VPN on AWS, GCP or Azure in about a minute. What it costs, how to verify it works, and the bug that cost me three instances.
↗Make a 30-second promo video — animation, voiceover, music, sound effects — with Claude and Claude Code. No paid APIs, no licensed music. Every prompt included.
↗Claude Code has four loop patterns: turn-based, goal-based (/goal), time-based (/loop and /schedule), and proactive routines. Which to use when, and how to make a loop actually finish.
↗Run Claude Code as the orchestrator and Antigravity's agy CLI as the worker so bulky legwork burns Google quota, not Claude tokens. Here's the split.
↗LLM API costs can kill a side project before it even launches. Here's how to pool OpenRouter, Gemini, NIM, and Groq using FreeLM to build a highly available, zero-cost AI stack for Python and Node.js.
↗Rate limits (HTTP 429) are the silent killers of AI applications. Learn how freelm implements intelligent circuit breaking and provider failover to achieve 99.9% uptime on free tiers.
↗The story behind freelm: frustration with dead API keys, the fragmentation of free AI models, and the journey to build a unified open-source gateway.
↗Don't rewrite your codebase to support multiple LLM providers. FreeLM provides a perfect OpenAI SDK shim that seamlessly routes traffic to Gemini, Groq, OpenRouter, and NIM.
↗Bootstrapping an AI startup? Don't pay for tokens until you have to. Learn how to combine multiple free API tiers into a reliable production stack using FreeLM.
↗Why and how I built freelm: an open-source Python client pooling six free-tier LLM providers behind one OpenAI-compatible call with failover and streaming.
↗Swap the OpenAI SDK for a free, drop-in replacement in one import. freelm.compat.OpenAI runs your existing chat.completions code on six free LLM providers.
↗Free LLM model IDs churn weekly. Instead of maintaining a list, freelm discovers them live from each provider's /models endpoint, caches and self-corrects.
↗A real walkthrough of publishing a Python package to PyPI in 2026: pyproject with hatchling, src layout, tests and trusted publishing — using freelm.
↗The $0 AI stack I use across side projects: six free-tier LLM providers pooled behind freelm, so prototypes run on free capacity with automatic failover.
↗I built LetX, a real-time collaborative LaTeX editor, using CRDTs for conflict-free editing and a sandboxed Docker compile pipeline. Here is the architecture.
↗QuantumSketch turns text prompts into narrated STEM animations using LLMs and ManimGL. Here is the pipeline that takes an idea to a finished video.
↗Vibe coding — writing software by describing intent to AI agents — is how I ship six products solo. Here's what it actually means, what works, and what breaks.
↗As a solo founder I run multiple products with AI coding agents, a shared skills layer and rigorous architecture. Here is the exact stack I use to ship.
↗BikroyBuddy is an AI shopping agent for Bangladesh that handles discovery and orders over chat, serving 5,000+ users. Here is how the architecture scaled.
↗Bagh is a Bangla-first programming language I designed that compiles to Python. Here is the compiler design and why a native language matters for coding education.
↗Cursor and Claude Code are the two dominant AI coding tools in 2026. Here's a direct comparison based on using both daily to ship six live products.
↗You can build global-class software from Dhaka. Here is what changes when your home is Bangladesh and your customers are global, and what gives you an edge.
↗I use Claude Code with a custom skills layer to plan, build, test and deploy across products. Here is my exact workflow and the skills I rely on.
↗Context-Heavy is a multi-tenant knowledge-graph API in Go (pgvector + recursive CTEs) that gives AI agents persistent context. Here is the architecture.
↗RAG (Retrieval-Augmented Generation) breaks in predictable ways at production scale. Here's the architecture that works based on running RAG in multiple live products.
↗CRDTs and OT both enable real-time co-editing; CRDTs merge conflict-free without a central server while OT transforms operations. Here is the breakdown.
↗GenAI video and agent workflows take minutes. Temporal.io makes them durable and retryable. Here is how QuantumSketch uses Temporal end-to-end.
↗pgvector, Pinecone, and Weaviate each have a real use case. Here's a direct comparison based on running vector search in production across multiple AI products.
↗I maintain open-source skills that supercharge AI coding agents. softco, skill-builder, openrouter-free-infer and more. Here is the full stack.
↗Microservices for one engineer pay off when you have the right tooling: Go services, Terraform IaC, AWS ECS Fargate. Here is the exact pattern I use.
↗Self-hosting open-source LLMs (Llama 3 70B, Mistral 7B) gives you privacy, cost control, and no rate limits. Here's the practical setup guide for 2026.
↗ChessGoddess analyzes games with Stockfish, explains moves with an LLM, and does image-to-FEN — deployed on AWS ECS. Here is the architecture.
↗GEO and AEO is still SEO with extra layers: answer-first content, dense structured data, and being in each AI engine's index. Here is the 2026 framework.
↗Go outperforms Python for AI backend services in concurrency, memory efficiency, and operational simplicity. Here's the case for Go-first AI backends in 2026.
↗ComiKola generates Bangla comics end-to-end — scripting, character design, panel images — on a Go + React + Temporal stack. Here is the architecture.
↗I build by owning the whole problem — design, code, ship, operate — not just implementing specs. Here are the principles that shape how I work.
↗common-knowledge is a Git-backed memory store I built so AI coding agents keep context across sessions. Here is the design and why Git is the right backend.
↗Next.js 15 App Router has a new Metadata API, React Server Components, and static export changes. Here's how to get maximum SEO and performance in 2026.
↗My BRAC University research used contrastive learning (SimCLR, Siamese nets) to detect exoplanets in astronomical imagery, reaching 92% accuracy. Here is the approach.
↗Run a private always-on AI agent on AWS for about $17/month using free OpenRouter models and Terraform. Here is the exact recipe and cost breakdown.
↗Multi-agent AI systems — where multiple LLM agents collaborate on complex tasks — require careful orchestration design. Here are the patterns that work in production.
↗I built a toolchain — skills, memory, free-LLM routing — that makes AI coding agents dramatically more capable. Here is the full set and what each enables.
↗offSchool builds personalized study plans with spaced repetition and knowledge graphs on a Next.js + Go + Python stack. Here is the architecture.
↗LLM API costs can spiral fast. Here are the techniques I use to keep costs under control across six AI products — from model tiering to prompt caching to batching.
↗After 170+ public and private repos, here is what building in public taught me about scope, momentum and reuse — and what I would do differently.
↗AI agents without memory forget everything. Here are the three memory patterns — RAG, episodic, and semantic — and when to use each in production agent systems.
↗Scaling a social-commerce AI agent to 300k+ users meant queueing, autoscaling on Kubernetes, and tight cost control. Here is the architecture that survived.
↗Deploying Next.js static exports to Cloudflare Pages gives you global CDN, zero cold starts, and free tier hosting. Here's the complete setup guide for 2026.
↗Prompt engineering in 2026 is less about tricks and more about structure. Here are the techniques that consistently improve LLM output quality based on real production use.
↗I've built six AI products solo in 2026. Here are the honest lessons — what the AI leverage myth gets wrong, what infrastructure decisions matter, and how to ship.
↗WebSockets in Go enable real-time features — live collaboration, chat, notifications. Here's how to build a production WebSocket server with proper connection management.
↗Go microservices in production require consistent patterns for HTTP, middleware, errors, config, and testing. Here are the patterns I use across six products.
↗pgvector adds vector similarity search to PostgreSQL. Here's the complete guide: installation, indexing (ivfflat vs HNSW), hybrid search, and production tuning.
↗Manim is a Python library for programmatic math animations. Here's how to create beautiful STEM visualizations — from basic shapes to 3D calculus animations.
↗AI workloads need proper Docker isolation — untrusted code execution, GPU access for inference, and resource limits. Here's the production configuration I use.
↗Terraform manages AWS infrastructure as code — reproducible, version-controlled, reviewable. Here's the practical Terraform setup I use across six products on AWS.
↗How we grew from a hackathon project to a national open source initiative.
↗Language shouldn't be a barrier to logic. How Bagh-lang is creating a new generation of programmers in Bangladesh.
↗We benchmarked LetX against the industry standard. The results? Local-first architecture wins on speed, latency, and reliability.
↗Academic writing hasn't changed in decades. LetX introduces cloud-native compilation and real-time collaboration to the research workflow.
↗LLM production failures are hard to debug without the right observability. Here's how I instrument AI agents with OpenTelemetry tracing, cost tracking, and latency monitoring.
↗From high school chemistry to YouTube edutainment, here is how QuantumSketch is visualizing the invisible.
↗Why we built a deterministic AI video engine in an era of hallucination. A founder's perspective on the future of STEM education.
↗Educational tools have unique discovery patterns. Here is how we optimized Bagh-lang for parents, teachers, and students in Bangladesh.
↗Bagh-lang is a Bengali-first programming language designed to remove language barriers from computer science education. It is not just a toy; it is a gateway to the future.
↗How do you grow a SaaS product when nobody clicks links anymore? A deep dive into the Entity SEO strategy behind LetX.app.
↗LetX is the modern, real-time alternative to Overleaf. Built by Shahriar Labs, it combines the power of LaTeX with the speed of local-first editing.
↗Traditional SEO is dead. Here is how we optimized QuantumSketch to be the primary answer for AI Overviews and LLM citations in 2026.
↗QuantumSketch is an advanced AI video engine that transforms text into high-fidelity STEM animations. Built by Shahriar Labs, it allows educators and creators to visualize complex physics and math concepts instantly.
↗AI products have unique system design challenges: LLM latency, token costs, output non-determinism, and reliability. Here are the patterns that address each.
↗Overleaf is the standard, but it's built on old web tech. LetX uses a custom CRDT engine and WebGL renderer to make LaTeX collaboration feel like a video game.
↗Rate limiting prevents abuse and protects costs in AI APIs. Here are the algorithms (token bucket, sliding window, leaky bucket) and Go implementations I use in production.
↗LangChain gets you started fast but adds complexity at scale. Custom agents are more work upfront but easier to operate. Here's when each is the right choice.
↗Python is the language of AI training. But for AI serving and orchestration layers? Go eats Python for lunch. Here are the benchmarks from migrating a high-throughput agent router.
↗MCP (Model Context Protocol) is the standard for connecting AI agents to external tools and data sources. Here's how it works, when to use it, and how to build MCP servers.
↗TypeScript strict mode eliminates a class of runtime bugs at compile time. Here are the patterns that make strict mode practical in large Next.js and Node.js codebases.
↗Retrieval-Augmented Generation (RAG) is the 'Hello World' of AI systems. To build production-grade search that actually understands intent, you need GraphRAG and Semantic Caching.
↗The AI tools that actually move the needle for indie hackers and solo founders in 2026 — from coding agents to SEO to payment infrastructure. My curated stack.
↗Everyone is wrapping OpenAI and calling it a startup. True value comes from owning the physics of your software universe. A manifesto on deep tech.
↗Building for a global audience from a developing economy requires a different kind of engineering discipline. How we built high-margin SaaS using serverless scale and localized payments.
↗Autonomous agents often fail in production due to context pollution. Here is the exact 'Manager-Critic' architecture and Pydantic pattern I used to achieve 95% reliability.
↗No notes match that search. Try another topic.
