SYSTEMS. AGENTS. DISTILLATION.
I'm Asim Ansari (@asimansari-git). I build small models, deterministic agentic workflows, and resilient backends that run close to the metal. Passionate about solving real-world problems through code—from low-level desktop CLI tools to human-in-the-loop multi-agent systems.
ENGINEERING ECOSYSTEM
NANOHAT OS AGENT
Sub-1B parameter Linux desktop agent distillation specially built for Fedora Workstation. Powered by a pure native Python harness, 250k verified action traces, and zero dependency bloat.
SAGE KNOWLEDGE VAULT
Private offline knowledge vault for deep engineering study. Dense ChromaDB vectors, SQLite FTS5 BM25, and Graphify AST graphs, exposed via a 9-tool Model Context Protocol server.
BILLPRODS AI
AI-powered item-level product intelligence ledger. Decomposes physical receipts into granular SKUs, tracks unit prices over time, and runs locally with FastAPI and Docker.
TEACHING LOOP
Agentic Socratic tutoring platform featuring human-in-the-loop intervention, multi-turn state tracing, and ChromaDB vector retrieval in a local Streamlit harness.
LIZ OS BRAIN
Personal autonomous workstation brain built on Google Antigravity SDK. Solves ADHD session amnesia through a Centralized Cortex (.liz/), OKF v0.1 specs, and custom skill execution.
TECHNICAL ARTICLES
Deep-dive technical articles on cyclic LangGraph state machines, AST tokenization traps, small model parameter efficiency, and modern backend systems.
OPEN SOURCE RUNTIMES & LIBRARIES
AI-powered item-level product intelligence & price history ledger. Decomposes physical receipts into granular SKUs, tracks unit price trends over time, and auto-categorizes items with Gemini AI & FastAPI.
An Agentic AI teaching platform featuring a human-in-the-loop architecture where teachers can dynamically guide AI tutors during student interactions with ChromaDB RAG.
Sub-1B OS desktop agent distillation harness built specifically for Linux (Fedora Workstation). Synthesizes 250k action traces with a pure native Python harness for deterministic system automation.
Progressive retrieval-augmented generation engineering curriculum evolving from basic similarity search into MMR reranking, metadata filtering, MultiQueryRetriever, and Self-Querying with ChromaDB and Google Gemini.
Experimental autonomous agent discovering its own runtime limitations and building self-solutions with no pre-built memory or hand-held guidance—just a persistent terminal harness and self-evolving loops.
CORE SPECIFICATIONS
01 // CYCLIC STATE MACHINES
Linear chains break under uncertainty. We design stateful cyclic graphs with LangGraph, featuring explicit state reducers and deterministic routing.
02 // SUB-1B OS DISTILLATION
Small models running locally without bloated dependencies. Distilling 250k synthetic action traces into lightweight models for native desktop tasks.
03 // ENTERPRISE BACKEND
Java 21 / Spring Boot 3 Virtual Threads and .NET 9 Native AOT Clean Architecture for high-throughput system stability and data integrity.
SELECTED ARTICLES
An architectural analysis of retry loops, state reducers, and error mitigation in autonomous workflows.
Synthetic multi-turn action trace curation, 4 mechanical validation gates, and standard library runtime harnesses.