Essays and notes on building, AI, music and everything in between.
48 postsAgents that act in the real world need supervision more than they need features. TETHER is an installable web app that supervises a real Fo run: one dispatch per submission, reported-versus-observed evidence, and a mascot that never pretends.
read →My captcha repo draws a six digit image, stores the answer in Redis and checks it in a verify endpoint. Reading main.py, a length check on an integer sits in front of the success branch.
read →My discovery-tiny-001 repo releases a small diffusion code model with a README that says it only knows 20 functions. Reading the inference script, there are two details worth knowing before you trust the 20 of 20.
read →My yetr-ai-agent repo is a Next.js front end for an MCP task runner. Reading the chat page, a failed send clears what you typed and the page leans on modules that are not in that folder.
read →My ai-browser-chat repo streams answers over server-sent events behind a login. Reading the chat page, the save step posts a stale copy of the conversation.
read →The analyze-url route in my ai-browser-chat repo hands a query-string URL to a data extraction call. Reading it, nothing limits which URLs it will accept.
read →My yeti-verse-connector-hub repo has a Netlify function that exchanges the LinkedIn code on the server, and a browser handler that holds the client secret. Reading both side by side.
read →My bonzai-core-engine repo routes tasks between E2B and Scrapybara by a complexity score. Reading the router, one of its routing rules matches text that nothing produces.
read →My weather-agent repo geocodes a city, fetches Open-Meteo data and asks a Groq model to plan activities. Reading the workflow, the high and low come from the wrong span of data.
read →snap-human-tools is a Python standard library server with three advisory tools. The interesting part of server.py is the rule layer that can overrule the model.
read →LocalFlow validates workflows with a schema table and a DFS cycle finder. Reading packages/core, one check is a placeholder, and the registry uses tool names the validator does not know.
read →The Modal service file in LifeGuide AI calls five health-related endpoints from the browser and falls back to canned answers when anything fails. Here is what that means for a wellness app.
read →ContextWeave AI's README lists WebGPU acceleration. I read the service and the panel. The retrieval core is real and small. The acceleration is a constant and a 1.5 second wait.
read →Neural Workspaces turns captures into a knowledge graph. The code extracts entities with a capital-letter regex and links every pair. Here is what that gives you and where the numbers it reports come from.
read →The MYND Platform README promises multi-agent orchestration. I read the Orchestrator class. It builds four agent roles and runs tasks one at a time, with a dependency wait that can never finish.
read →MYND Plan is my deployable agent app. The chat route runs a real tool loop capped at six hops, with retries, a timeout and a rate limit. Here is what the code does and what it does not.
read →The README for MYND Synchron mentions CRDT operations. The sync engine in the code hashes state, encrypts it and merges by diffing. That is a different thing, and I explain why the label matters.
read →My Knowledge Paleontology repo borrows a geology metaphor for tracking where knowledge came from. The code underneath is a SHA-256 Merkle tree with proofs. I look at what it proves and what it leaves open.
read →Prompt Phylogeny draws prompt families as an evolutionary tree. The fingerprinting is real code. The clustering and the fitness scores are placeholders, and the file says so.
read →MYND Skill Forge describes a platform for authoring, testing and distributing agent skills. The repository holds a README and a plan, and I say so up front.
read →Knowledge distillation was described by Hinton, Vinyals and Dean in 2015. MYND Model Holography is a visualization project around that idea, built with Three.js and Python workers.
read →MYND Decision Paleontology records the lineage of an agent's decisions. EU AI Act Article 12 requires logging for high-risk systems. They are related, and they are not the same thing.
read →NeuralShield-AI stacks seven defense layers around an LLM. The README opens with the sentence I care about most: no monitor can guarantee full detection.
read →QuantumCrypt-AI implements ML-KEM, ML-DSA and SLH-DSA in Python with no C dependencies. The README says plainly that it has not had a third-party audit.
read →VaultFlow keeps a freelancer's contracts, payment milestones and shared files on infrastructure they control, with SQLite and Docker.
read →MYND Model Arena runs models against each other in tournament battles and ranks them with ELO. Here is what it measures and where the idea comes from.
read →KeepTrak keeps a solo founder's invoices, expenses and contracts in a local SQLite file. No account, no cloud, no telemetry.
read →RBSP implements BM25, TF-IDF, HNSW vector search and a Porter stemmer in pure Python. The dependencies field in pyproject.toml is empty.
read →MYND Agent Replay records every LLM call, tool invocation and state change an agent makes, then lets you replay the session frame by frame.
read →KARMA builds causal graphs and runs counterfactuals with DoWhy and EconML, so a decision tool answers what will happen if you act, not just what moved together.
read →MuleShield AI maps its features to RBI, FIU-IND, PMLA and DPDP requirements before it talks about models.
read →Feen stores provider keys encrypted and hands teammates limited access tokens instead of the keys themselves.
read →Project Anvil is a self-hosted set of Drive, Docs, Video, Maps, Search and Mail apps that share one sign-in.
read →The Quantum Lotus came out in 2025: 49 pages, set in a world remaking itself after environmental collapse.
read →A debut single in 2024, ten more in 2025, in Malayalam and Hindi. What the discography looks like when you list it plainly.
read →I shoot pistol, 10 m and 25 m, and I won state gold in 2024. What a sport that rewards stillness has in common with debugging.
read →Palaka Lite is a Rust agent harness. Its central rule: a prompt injection can make the model ask for anything, but the runtime decides what executes.
read →Mynd Domains hands out yourname.mynd.my.id subdomains. The only way to get one is a merged pull request in a public repository.
read →I ternarized Intern-S2 after training and shipped the result as 24 files. What TQ2_0 stores, what the byte counts say, and what they do not.
read →GitHub release assets must be under 2 GiB. The Intern-S2 builds are far larger. The arithmetic of splitting them, and what I changed between versions.
read →Ezhuthaani is a port of Visual Studio Code that replaces Electron with Tauri. What changes, what stays, and a 12 MB Windows installer.
read →Varsha encrypts documents on your device. The hard part is not the cipher. It is being honest about ownership, recovery, and backup.
read →LevelCode separates finding files, planning changes, editing, and review. More agents help only when their responsibilities are clear.
read →Discovery-Tiny-001 reproduces 20 training functions exactly. That is a narrow result worth keeping narrow, with code and evaluation attached.
read →Most personal sites are one dead CMS away from a blank page. Mine treats the CMS as optional. The architecture behind founder.myndlabs.tech, and why a name deserves infrastructure that cannot fail.
read →The counter on my homepage says 180 original projects. There is no team behind it. What building in public, alone, from Kerala actually looks like, and what the reps teach you.
read →INDU is my answer to a question every serious system keeps asking: why choose between a language that is pleasant, one that survives failure, and one that is fast?
read →A reflective essay on building software with AI coding assistants — what happens when a session with tools like GitHub Copilot, Claude Code and Devin turns into something unexpectedly human.
read on medium.com →