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 (Qwen3-0.6B) purpose-built for Fedora Workstation. Powered by a hardened 10-tool direct-dispatch Python harness, sub-700ms CPU latency, and zero cloud dependencies.
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 on-device OS desktop agent running natively on Fedora Workstation. Aligned on Qwen3-0.6B with a hardened 10-tool direct-dispatch runtime for sub-second system telemetry and daemon control.
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 LOCAL TOOL CALLING
Compact language models running locally on CPU without heavyweight agent frameworks. Pairing specialized 0.6B tool-calling weights with zero-shell Python execution harnesses for native Linux control.
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.
Orthogonal 10-tool schema contracts, zero-shell Python dispatch, Wayland session crash guards, and sub-second local CPU execution on Fedora Workstation.