Phillip Johnson

System Administrator

SeniorAI/ML Engineer

2years in support

6/10Data Pipeline & Feature Engineering

3years in tech

Overview

A pragmatic LLM-focused engineer (mid-career) who builds end-to-end local LLM pipelines and image-to-ASCII content tooling; the strongest proven skill is engineering production-ready orchestration for local LLM analysis and embedding-based deduplication as shown by News-digest/src/pipeline.py and News-digest/src/dedup.py; the public code does not show custom model training, experiment tracking, or GPU/serving optimizations.

Phone

Timeline

NOC Engineer · Middle

Teleport · Full-Time

Mar 2025 to Sep 2026 1 Year 6 Months Izhevsk In office

Zabbix

TCP/IP

DNS

Wazuh

Linux

IT Support (unofficial practical experience) · Middle

Small Business · Contractor

Jan 2024 to Aug 2026 2 Years 7 Months Izhevsk Partially remote

Provided hands-on IT support tasks for a small business environment. Commissioned an HP blade server setup and connected IP cameras to server storage using media streaming tools. Set up remote connectivity between branch locations and configured authenticated file sharing for users. Focused on getting Linux/storage services and access control working reliably.

Linux

Ubuntu

VPN

Wi-Fi

Projects

Internet Mirror

Mar 2026 to Jun 2026 3 Months

Built a local infrastructure to mirror and distribute software packages and services. Used Proxmox with virtual machines and containers to host mirrors, including Nextcloud and package repositories. Configured MikroTik networking components and set up multiple WireGuard instances for connectivity between networks.

Proxmox VE

Linux

Ubuntu

VPN

Bash

Offline AI Translator Pipeline

Designed and validated an offline speech translation pipeline optimized for low end-to-end latency. Integrated a speech processing and translation workflow using Whisper for transcription and a translation model for output. Selected an embedded platform concept with an NPU to support the pipeline requirements.

Virtualization and Windows Infrastructure Lab

Created a practice lab for Active Directory and Windows domain concepts. Deployed Proxmox with Windows Server and Windows clients to simulate domain structures and domain-join scenarios. Used the environment to validate configuration steps before applying ideas to real deployments.

Technical skills

Languages

Python

Bash

Other

Requests

LLM

NumPy

OpenSSH

Nmap

FFmpeg

Claude

Embeddings

Active Directory

GitHub

Wazuh

Linux

VMWare

Windows

Docker Compose

Ubuntu

Containers

Debian

Proxmox VE

TCP/IP

Windows Server

KVM

DHCP

DNS

Gitea

Senior AI/ML Engineer Confidence: Medium LLM Engineer

Model Architecture & Training

2/10

How well models are designed and trained

Minimal model-building; code shows structured use of hosted/local LLMs and embeddings with fallbacks but no custom training or architectures.

Evidence

News-digest/src/pipeline.py: call_ollama_json implements model selection, fallback and JSON-enforcing format

News-digest/src/dedup.py: OllamaEmbedder.embed and embed_many call Ollama embeddings endpoint

Data Pipeline & Feature Engineering

6/10

How data is prepared for models

Solid data pipelines and feature handling for news ingestion and image mask processing; includes RSS/Search collection, extraction, embedding-based deduplication and persistent seen store plus careful image mask construction and sampling.

Evidence

News-digest/src/search.py: collect_candidates, fetch_rss, search_searxng - gathers and filters RSS and SearXNG results

News-digest/src/dedup.py: SeenStore uses chromadb.PersistentClient and cluster_by_similarity implements greedy clustering

ASCII-Wallpaper/tools/gen_scene.py: functions cm, tensor, mask handling and rendering pipeline that transform image + PSD masks -> ASCII scene

Experimentation & Evaluation

1/10

How results are measured and tested

Minimal experimentation/evaluation artifacts; structured prompts and logging exist but there is no experiment tracking, metrics collection or reproducible experiment harness.

Evidence

News-digest/src/pipeline.py: FACTS_SYSTEM/VERIFY_SYSTEM templates and JSON enforcement for model outputs

News-digest/src/main.py: logging and orchestration but no experiment tracking (W&B/MLflow) or eval harness

MLOps & Deployment

4/10

How models are shipped to production

Basic MLOps aspects: persistent embedding store, local LLM orchestration with timeouts and fallback attempts, and an end-to-end orchestrator that writes reports; lacks production-grade serving, monitoring or versioned model deployment.

Evidence

News-digest/src/dedup.py: SeenStore uses chromadb.PersistentClient and a get_or_create_collection pattern

News-digest/src/pipeline.py: call_ollama_json implements retries and fallback model logic and uses request timeouts

Computational Efficiency

3/10

How efficiently computing resources are used

Some attention to computational efficiency in image processing via NumPy vectorization and limiting / downsampling strategies, but no GPU/quantization/distributed optimizations for model inference or embeddings.

Evidence

ASCII-Wallpaper/tools/gen_scene.py: vectorized numpy operations, kmeans implementation, subsampling (SS_X, SS_Y) and limits (surfaces)

ASCII-Wallpaper/tools/make_draft_psd.py: use of scipy.ndimage filters and image-resize strategy for mask generation

Research Depth & Innovation

2/10

Depth of research and new ideas

Original heuristics and practical algorithms are present (image-to-ASCII heuristics, stream ordering), but no research-level contributions or reproduced paper results.

Evidence

ASCII-Wallpaper/tools/gen_scene.py: make_stream heuristic that orders stream cells with perspective correction

ASCII-Wallpaper/tools/compare_styles.py: fs_dither, custom tone/edge heuristics and Braille/block encoding

Expertise

LLM· Middle

Industries

Cybersecurity· Senior

Media & Entertainment· Middle

Technologies

VMWare

Debian

Embeddings

Windows Server

NumPy

Ubuntu

LLM

Gitea

Proxmox VE

KVM

GitHub

Linux

Windows

TCP/IP

DNS

DHCP

Recommendations

Develop LLM-based ingest and summarization/verification workflows (local Ollama or self-hosted LLMs) with robust embedding dedup and persistent stores
Build content pipelines that combine image preprocessing and procedural generation (image->mask->ASCII renderer) for UI or creative tooling
Extend existing tooling with experiment tracking and evaluation (W&B/MLflow) and add unit/integration tests for the extraction and dedup pipelines
Harden MLOps: add model/versioning, monitoring and batched embedding inference to improve efficiency and operational reliability

Repositories

The developer's experience in this domain has been verified based on AI analysis of the following repositories:

ASCII-Wallpaper News-digest

Middle Backend Developer Confidence: Medium Data Platform

A middle-level backend developer focused on pragmatic Python pipelines for ETL and LLM/embedding orchestration. The strongest proven skill is designing and implementing an LLM-driven news ingestion and deduplication pipeline as evidenced by News-digest/src/pipeline.py, News-digest/src/dedup.py and News-digest/src/main.py. There is little evidence of production-grade infra such as CI/CD, service-level testing, formal migrations, secrets management, or load-testing artifacts in public code.

API Design

3/10

How well APIs are designed

Basic external API integration with timeouts and a fallback strategy, but no formal API versioning, idempotency handling or consistent error contract for downstream callers.

Evidence

News-digest/src/pipeline.py: call_ollama_json - requests.post with timeout and fallback model retry logic

News-digest/src/search.py: search_searxng - requests.get to SearXNG with timeout and params

Data Layer & Database

3/10

Working with databases

Practical use of a vector store (ChromaDB) and in-memory clustering for deduplication, but no schema migrations, transactions, or advanced query tuning are present.

Evidence

News-digest/src/dedup.py: SeenStore - chromadb.PersistentClient usage, collection.query and add

News-digest/src/dedup.py: cluster_by_similarity - greedy cosine-sim clustering implementation

Scalability & Performance

2/10

Handling load and speed

Some attention to performance in hot paths (vectorized numpy image processing and batching-like embedding calls) and rate controls, but no queueing, measured load-testing, or clear caching/invalidation strategy for high-throughput production workloads.

Evidence

ASCII-Wallpaper/tools/gen_scene.py: heavy use of NumPy vectorized operations and kmeans for image -> scene conversion

News-digest/src/main.py: orchestration with TOP_ITEMS_PER_TOPIC and search.collect_candidates uses request_delay_seconds to throttle requests

System Architecture

4/10

Overall system structure

Clear modular decomposition into search, extract, dedup, pipeline and report stages with a single orchestrator, showing deliberate pipeline separation and replacement points for components.

Evidence

News-digest/src/main.py: process_topic orchestrates collect_candidates, extract, dedup, cluster, cross-verify and report

News-digest/src/pipeline.py: explicit multi-step functions extract_facts, cross_verify, daily_thesis separating concerns

Security & Auth

2/10

Protecting data and access

Some security-aware choices such as request timeouts and local LLM hosting preference, but little evidence of auth/token lifecycle management, secrets handling, input hardening, or dependency vulnerability auditing.

Evidence

News-digest/src/pipeline.py: call_ollama_json uses request timeout and fallback model selection

News-digest/src/search.py: fetch_rss and search_searxng wrap requests in try/except and use timeouts

Reliability & Observability

3/10

Stability and monitoring

Reasonable logging and defensive exception handling across the pipeline with fallback logic, but missing structured tracing, metrics, graceful shutdown, and systematic retry/backoff policies.

Evidence

News-digest/src/main.py: logging.basicConfig and try/except around per-topic processing to continue on failure

News-digest/src/pipeline.py: call_ollama_json logs warnings and tries fallback model before raising OllamaError

Expertise

Backend AI & LLM· Middle

Databases & Vector Storage· Middle

Python· Middle

Recommendations

Develop embedding/LLM-driven data pipelines and local LLM orchestration (embedding calls, dedup, cluster workflows).
Implement small to medium ETL systems that ingest, normalize and deduplicate text feeds with ChromaDB-like vector stores and offline report generation.
Build image-processing utilities and content generators that leverage NumPy/Pillow for deterministic previews and offline asset generation.

Repositories

The developer's experience in this domain has been verified based on AI analysis of the following repositories:

ASCII-Wallpaper News-digest

Middle DevOps Engineer Confidence: Medium Generalist

Generalist developer at a middle level (tier_score 3.0) with a strong practical focus on modular Python tooling for image processing and local LLM-based pipelines. The strongest proven skill is building modular Python pipelines that orchestrate local LLM calls, embedding/deduplication, and HTML report rendering as implemented in News-digest (src/main.py and src/pipeline.py). There is little evidence of production-grade CI/CD, infrastructure-as-code, Kubernetes orchestration, or observability/SLO engineering in public code.

CI/CD Pipelines

1/10

Automated build and deploy

Minimal evidence of CI/CD; only a simple docker-compose service exists and no reusable workflows, gating, or artifact signing.

Evidence

News-digest/docker-compose.yml: searxng service provided as a compose entry (image: searxng/searxng:latest, ports/volumes configured)

Infrastructure as Code

Managing servers with code

Not evidenced in public code

Containerization & Orchestration

1/10

Working with containers

Basic container usage via docker-compose is present but no orchestration (Kubernetes), resource tuning, multi-stage builds, or production-grade pod settings are present.

Evidence

News-digest/docker-compose.yml: docker-compose usage for searxng service

Observability & Monitoring

1/10

Watching system health

Only basic programmatic logging is used; there are no SLOs, burn-rate alerts, structured tracing, dashboard-as-code, or alert-routing artifacts.

Evidence

News-digest/src/main.py: logging.basicConfig(...) and logger usage

News-digest/src/pipeline.py: log = logging.getLogger("pipeline") and log.warning/error calls

Reliability & Incident Response

2/10

Keeping systems up

Application-level reliability practices exist (fallback model attempts, per-topic try/except, error logging) but there is no evidence of deploy strategies, runbooks, postmortems, chaos testing, or automated rollback gates.

Evidence

News-digest/src/pipeline.py: call_ollama_json implements model fallback loop with exception handling and logging

News-digest/src/main.py: process_topic wrapped in try/except to skip failed topics and log.exception

Cloud & Cost Optimization

1/10

Smart use of the cloud

Some cloud/resource-awareness in config (model sizes, local Ollama host) exists but no autoscaling, spot/eviction strategies, IAM, or cost-measurement artifacts.

Evidence

News-digest/config/topics.yaml: llm.analysis_model, fallback_model, and comments about VRAM/RAM requirements

Industries

Artificial Intelligence· Middle

Technologies

Containers

Python· since 2023 · Senior

Docker Compose

Requests

Recommendations

Develop and harden local LLM orchestration and embedding services into a reproducible deployment with CI pipelines and container images (e.g., add multi-stage Dockerfiles and reusable CI jobs).
Expand operational posture for services: add container orchestration or deployment manifests, runtime resource limits, health probes, and a basic alerting/SLO strategy integrated with logging/metrics.
Package the image-processing tools as versioned modules or CLIs and add automated tests and example workflows to increase reusability and onboarding.

Repositories

The developer's experience in this domain has been verified based on AI analysis of the following repositories:

ASCII-Wallpaper News-digest

Job preferences

Work permits: Georgia, Russia
Citizenship: Russia
Employment type: Full-Time

Open to AI agents: they read this profile and start a dialog through Alionalion.io/@phillip-johnson/mcp