# Asim Ansari — Full Engineering Specification & System Curriculum

> Autonomous Systems Engineer, Small Language Models (SLMs) Distillation Specialist, and Backend Infrastructure Engineer.

---

## 1. Professional Overview & Engineering Philosophy

- **Name:** Asim Ansari
- **Online Identity:** @asimansari-git (GitHub), link-asimansari (LinkedIn), @asimibnakhlaque (X/Twitter)
- **Email:** asimibnakhlaque@gmail.com
- **Core Specialization:** Autonomous agent runtimes, sub-1B parameter model distillation for deterministic desktop action execution, hybrid retrieval-augmented generation (dense vector + sparse BM25 + AST graph), and enterprise backend infrastructure (Java 21/Spring Boot 3, Python/FastAPI, C#/.NET 9).
- **Academic Background:** Master of Computer Applications (MCA), ICFAI University.
- **Engineering Philosophy:**
  1. *First-Principles Verification:* Rejection of shallow wrapper libraries; deep understanding of state transitions, vector embeddings, and memory hierarchies.
  2. *Deterministic Execution:* Prioritizing cyclic reducer graphs and validation loops over brittle linear LLM chains.
  3. *Local-First & Resource Efficiency:* Distilling capabilities into compact models (SmolLM2-360M, Qwen2.5-0.5B) capable of CPU-only execution without telemetry leakage or cloud latency.
  4. *Interface Engineering:* Strict monochrome palette, high typographic contrast, and zero layout shift over superficial decorative telemetry.

---

## 2. Core Project Architectures & Technical Blueprints

### A. NanoHat OS Agent (Flagship Autonomous System)
- **Summary:** Sub-1B parameter local desktop automation runtime tailored for Linux/Fedora Workstation environments.
- **Model Hierarchy:** Distilled student checkpoints (SmolLM2-360M, Qwen2.5-0.5B) trained on 250,000 verified multi-turn POSIX desktop action traces via QLoRA.
- **Harness & Tool Routing:** 100% native Python runtime without LangChain/CrewAI wrappers. Interacts directly with Linux OS subsystems:
  - Process Inspection: `/proc` filesystem parsing, PID kill/signal handling, `systemd` unit control.
  - POSIX Shell Execution: Sandboxed subshell invocation with timeout enforcement and stderr capture.
  - Native GUI & Window Management: `wmctrl`, `xdotool`, and GNOME D-Bus IPC for window focus and virtual keystroke dispatch.
  - Network Diagnostics: ICMP socket ping, DNS resolution, and interface routing table audits.
- **Key Metrics:** <85ms local inference latency on modern CPU, zero telemetry leakage, 98.4% single-step command execution accuracy on benchmark suites.
- **Repository:** https://github.com/asimansari-git/NanoHat
- **Case Study:** https://asimansari.com/projects/nanohat.html
- **Raw Markdown:** https://asimansari.com/projects/nanohat.md

### B. Sage Knowledge Vault (Internal Knowledge Engine)
- **Summary:** Private, air-gapped knowledge retrieval and reasoning vault operating over an internal corpus.
- **Architecture:** Hybrid multi-index retrieval system:
  - Dense Vector Store: ChromaDB with cosine distance indexing over chunked markdown and technical specifications.
  - Sparse Lexical Index: SQLite FTS5 with BM25 ranking for exact keyword, function name, and identifier matching.
  - Knowledge Graph AST: Code AST parsing via Graphify to model semantic edges and dependency hierarchies.
  - Security & Sandbox: Strict Torikago containerization preventing network egress and unauthorized filesystem traversal.
- **Interface:** Exposed locally through a 9-tool Model Context Protocol (MCP) server integration.
- **Case Study:** https://asimansari.com/projects/sage.html
- **Raw Markdown:** https://asimansari.com/projects/sage.md

### C. BillProds AI (Product Intelligence & Vision Ledger)
- **Summary:** Item-level grocery and retail intelligence ledger transforming physical paper receipts into structured financial time-series.
- **Architecture:**
  - Ingestion & Vision: Asynchronous OCR pipeline using Gemini Vision and fine-tuned boundary detectors to extract store, timestamps, SKU items, unit prices, and category tags.
  - Deduplication & Reconciliation: Levenshtein distance and embedding similarity to unify fragmented store item names into global product entities.
  - Backend Stack: Python 3.12, FastAPI, PostgreSQL / SQLite with SQLAlchemy, Pydantic v2 schemas.
  - Frontend: Progressive Web App (PWA) with responsive mobile client and offline caching.
- **Repository:** https://github.com/asimansari-git/billprods-ai
- **Case Study:** https://asimansari.com/projects/billprods-ai.html
- **Raw Markdown:** https://asimansari.com/projects/billprods-ai.md

### D. Teaching Loop (Agentic Socratic Tutoring Engine)
- **Summary:** Interactive educational platform enforcing active recall and Socratic guidance with automated Human-In-The-Loop (HITL) teacher escalation.
- **Architecture:**
  - Cyclic State Machine: Manages dialogue state, tracks learner misconception markers, and gates hint delivery.
  - Retrieval Layer: ChromaDB indexing official syllabus documents, lecture transcripts, and rubric guides.
  - Teacher Intervention Protocol: Triggers asynchronous alerts to human educators when student frustration scores or repeated failure loops exceed thresholds.
  - Stack: FastAPI backend, ChromaDB vector store, multi-page Streamlit instructor dashboard.
- **Repository:** https://github.com/asimansari-git/teaching-loop
- **Case Study:** https://asimansari.com/projects/teaching-loop.html
- **Raw Markdown:** https://asimansari.com/projects/teaching-loop.md

### E. Liz OS (Workstation Orchestration Engine)
- **Summary:** Personal development assistant and context orchestration runtime operating locally on Linux.
- **Architecture:**
  - Runtime: Agentic sidecar architecture utilizing native tool dispatch and local memory stores.
  - Memory System: Markdown-based Open Knowledge Format (OKF v0.1) maintaining 3-day operational sliding windows, skill documentation, and project registries.
  - Subagent Coordination: Dispatches specialized subagents for research, code editing, and background verification tasks.
- **Case Study:** https://asimansari.com/projects/brain-liz.html
- **Raw Markdown:** https://asimansari.com/projects/brain-liz.md

### F. Gemini-RAG (Progressive Retrieval Curriculum)
- **Summary:** 6-stage algorithmic benchmark curriculum isolating modern retrieval techniques on isolated Git branches.
- **Stages:**
  1. *Naive Vector Search:* Base similarity lookup exposing chunking boundary degradation and query-document drift.
  2. *Recursive Character Chunking:* Semantic boundary preservation with overlap tuning.
  3. *Maximal Marginal Relevance (MMR):* Algorithmic balancing of relevance and diversity to prevent redundant context retrieval.
  4. *SelfQueryRetriever:* Dynamic metadata filter extraction from natural language questions using LLM query translators.
  5. *Contextual Compression & Re-Ranking:* Post-retrieval context pruning with Cross-Encoder rankers.
  6. *Hybrid Search & Fusion:* Reciprocal Rank Fusion (RRF) combining dense embeddings and sparse BM25 scores.
- **Repository:** https://github.com/asimansari-git/Gemini-RAG
- **Case Study:** https://asimansari.com/projects/gemini-rag.html
- **Raw Markdown:** https://asimansari.com/projects/gemini-rag.md

---

## 3. Technical Articles & Deep Dives

### Article 1: Why Linear Prompt Chains Fail: The Case for Cyclic StateGraphs
- **URL:** https://asimansari.com/blogs/cyclic-stategraphs.html
- **Raw Markdown:** https://asimansari.com/blogs/cyclic-stategraphs.md
- **Key Concepts:**
  - *Failure Modes of DAGs:* Directed Acyclic Graphs cannot recover from transient tool failure, invalid JSON outputs, or hallucinated arguments without catastrophic termination.
  - *Cyclic Reducer Architecture:* State transition machines where agents evaluate feedback, maintain immutable state histories, and cycle back to remediation nodes.
  - *LangGraph State Machines:* Practical implementation patterns using TypedDict state schemas, conditional edges, and persistent checkpointers.

### Article 2: Distilling Desktop Automation into Sub-1B Models
- **URL:** https://asimansari.com/blogs/sub-1b-desktop-distillation.html
- **Raw Markdown:** https://asimansari.com/blogs/sub-1b-desktop-distillation.md
- **Key Concepts:**
  - *The Latency & Privacy Bottleneck:* Frontier models (GPT-4, Claude 3.5) introduce 1-3s round-trip latency and risk confidential clipboard/file data exposure.
  - *Distillation Pipeline:* Synthesizing multi-turn terminal and desktop interaction datasets using frontier teacher models, followed by LoRA/QLoRA fine-tuning on SmolLM2-360M and Qwen2.5-0.5B.
  - *Grammar-Constrained Decoding:* Enforcing JSON schema validity at the token logit level during local CPU inference.

---

## 4. Technical Stack & Tooling Matrix

- **Languages:** Python 3.12+, Java 21 (LTS), C# 13 / .NET 9, SQL, Bash / POSIX Shell.
- **AI & Agentic Frameworks:** LangChain, LangGraph, Google Antigravity SDK, Ollama, Llama.cpp, ChromaDB, Hugging Face Transformers & PEFT.
- **Backend Infrastructure:** FastAPI, Uvicorn, Spring Boot 3, ASP.NET Core Minimal APIs, PostgreSQL, SQLite (with FTS5), Docker, Podman.
- **Operating Systems & Environments:** Fedora Linux Workstation, Linux POSIX subsystems, Systemd, D-Bus, X11/Wayland automation tooling.
- **Architecture Paradigms:** Event-driven architecture, Cyclic State Graphs, Microservices, Clean/Hexagonal Architecture, Test-Driven Development.

---

## 5. Contact & Social Graph

- **Portfolio Website:** https://asimansari.com
- **GitHub Profile:** https://github.com/asimansari-git
- **LinkedIn Profile:** https://linkedin.com/in/link-asimansari
- **X / Twitter:** https://twitter.com/asimibnakhlaque
- **Direct Email:** asimibnakhlaque@gmail.com