Farhad Agzamov

About

I build production LLM systems on Claude. At Fliplet, that's meant three things:

KnowledgeMCP — a Python MCP server on AWS Fargate that lets engineers navigate our 40+ repositories in natural language.
An adaptive PR review bot that picks between Claude models based on how much reasoning the change actually needs.
A library of 20+ Claude Code skills that collapse daily engineering workflows into single commands.

Before Fliplet, 14 years leading teams on museum-scale installations in Abu Dhabi, AWS serverless platforms, and WebGL campaigns for major film releases.

My favourite work happens at a whiteboard with a founding engineer, figuring out how their system should actually work.

Core Competencies

Applied AI & LLMs

Claude API (Sonnet 4.5 / Opus 4.6) Claude Code (Skills, Sub-agents, Hooks) Model Context Protocol (MCP) RAG & Vector Search (ChromaDB) Embeddings (nomic-embed-text, sentence-transformers) Context Engineering Prompt Caching Tool Use & Structured Output Adaptive Model Selection LangChain

Engineering & Infra

Python Node.js Vue.js / Nuxt.js React / Next.js FastMCP / Uvicorn AWS (ECS Fargate, Aurora, Lambda) Docker GitHub Actions / CI/CD OAuth2 / JWT / PKCE Datadog / Sentry Three.js / WebGL / Unity

Customer & Delivery

Technical Discovery Architecture Advisory Stakeholder & Exec Communication Cross-team Coordination On-site Deployment & Workshops Incident Command & RCA Budget & Resource Management

Domain Expertise

Developer Platforms & Internal Tooling Museum-scale Digital Experiences Interactive Installations IoT & Sensor Integration AR/VR & Projection Mapping Hardware Integration

AI Systems Shipped

Scale & Deployment

40+ repositories and 5 documentation sources indexed into 60,890 documents across 15 ChromaDB collections
Available to the Fliplet engineering team via REST API and native Claude Code integration
Deployed on AWS ECS Fargate (eu-west-1, 0.5 vCPU / 2 GB) at kb-mcp.fliplet.com with auto-deploy on every push to main

Engineering Decisions

Hybrid embedding strategy — nomic-embed-text-v1.5 (768-dim) via sentence-transformers in prod, Ollama locally, switched by env var with zero code change
Incremental re-indexing with Git SHA tracking (7-day rolling window) and adaptive batch sizes (10/20/50) to avoid OOM on large monorepos
OAuth2 + JWT + PKCE auth flow with email-domain validation for Claude Code native MCP handshake
Pre-computed ChromaDB committed via Git LFS — new engineer onboarding drops from 2–3 hrs to 9 min
Freshness metadata (last_modified, ingested_at, age warnings) surfaced inline with every search result

Impact: Engineers query architecture, APIs, conventions, and patterns across the entire estate in natural language. Weekly incremental re-index (GitHub Actions) keeps the index fresh with minimal compute. Five MCP tools exposed: search_fliplet_docs, search_fliplet_code, check_fliplet_freshness, list_fliplet_sources, search_customer_docs.

Python 3.11 FastMCP ChromaDB LangChain nomic-embed-text AWS ECS Fargate Docker OAuth2 / JWT / PKCE GitHub Actions

Key Mechanisms

Two-pass triage — first pass flags candidate files via Sonnet 4.5 (fast, cheap); second pass reviews only flagged files, escalating to Opus 4.6 when a 0–100 complexity score crosses threshold
Claude tool_use for guaranteed JSON-schema output — no free-text parsing, no hallucinated format drift
Ephemeral prompt caching on the (fixed) system prompt and coding standards block — materially reduces latency and cost on frequent-PR repos
Severity-based routing — findings bucket into block / warn / nit so reviewers see signal, not noise

Engineering Details

3,246 lines of Node.js across the GitHub Action source tree (claude-reviewer.js, complexity-scorer.js, constants.js)
Model IDs pinned per repo — claude-sonnet-4-5-20250929 default, claude-opus-4-6 escalation path
Runs as first-class GitHub Action across Fliplet repositories; comment threads attach directly to PR diff lines
Per-repo coding standards are cacheable inputs — standards change rarely, PRs change constantly, caching is the right call

Impact: Reviews every PR automatically with consistent standards, freeing senior engineers for architecture and design discussions. Model selection is measurable — not "always use the biggest model" — which keeps inference spend aligned with value per review.

Claude Sonnet 4.5 Claude Opus 4.6 Tool Use / Structured Output Prompt Caching Node.js GitHub Actions Anthropic SDK

Scale & Adoption

20+ user-authored Claude Code skills across personal (~/.claude/skills/) and team-distributed (ai-tool-library/skills/fliplet-engineering/) libraries
Skills chain together GitHub, Sentry, Datadog, Jira, AWS, Google Workspace, and custom MCPs — a working blueprint for "Claude Code as an engineering operations surface"
Team-distributed subset (12 skills) packaged for Fliplet engineering: PR review, widget audits, incident investigation, deploy readiness, platform health

Skill Patterns Shipped

Standards enforcement — fliplet-pr-review, widget-check, full-widget-audit (Vue 3 composition API, security, accessibility, Sentry integration)
Incident & production surface — investigate, check-platform, check-sentry-errors, post-deploy-check (chain Jira + Sentry + Datadog + AWS MCPs)
Daily rotation & context — gardener, my-prs, morning-briefing (keep engineers on the right signal, not the loudest one)
Meeting & comms context — prep-meeting, research-attendee, post-meeting, internal-comms (Gmail + Calendar + Drive + transcripts into briefs)
Scaffolding & delivery — fliplet-widget-setup, deploy, update-apps

Why this matters for an Applied AI SA / FDE role: Every skill is a case study in translating a team workflow into an LLM-native tool — the exact conversation a founding engineer needs to have when adopting Claude. First-hand fluency with Anthropic's extensibility stack (sub-agents, hooks, slash commands, MCP integration) means I can pattern-match a startup's workflow to a working skill or agent in minutes, not weeks.

Claude Code Sub-agents Slash Commands Hooks MCP Integration Anthropic SDK Skill Authoring

Experience

Lead engineering on a no-code/low-code platform serving 100,000+ monthly users across 58,000+ apps in 3 AWS regions — Node.js, Vue.js, AWS Fargate, Aurora PostgreSQL
Shipped three production AI systems on Claude — KnowledgeMCP (custom MCP server), an adaptive PR review bot, and a 20+ skill Claude Code library. See the AI Systems section above for architecture and engineering decisions
Drove 60–70% reduction in alert noise across 110+ Datadog monitors (downgraded 22, removed 4 broken, added production-env filters to 17) and replaced Datadog relay with 7 direct Sentry rules; caught 4 silent dead monitors via chaos testing
Led incident command and authored 16 P0/P1 RCAs including a 2-hour cross-region SSO/SAML outage (root cause: SameSite cookie regression), with contributing-factor analysis and preventive measures adopted into release process

Claude API MCP / FastMCP ChromaDB LangChain Python Node.js Vue.js AWS ECS Fargate Aurora PostgreSQL GitHub Actions Datadog Sentry

Led technical delivery for museum-scale digital experiences across the National History Museum Abu Dhabi, Sheikh Zayed Grand Mosque Visitor Center, Early Childhood Authority, and Arabic Language Museum — leading teams of up to 20 and liaising with 50+ stakeholders across client, hardware, and creative partners
Architected web solutions for high-traffic museum environments with multilingual digital labels, interactive 3D displays, and hardware-integrated exhibit control systems
Introduced DevOps and automated testing practices to a previously manual delivery pipeline, reducing deployment failures and release cycle time

Vue.js Nuxt.js React Next.js Three.js Node.js AWS Azure Docker

Built and led the engineering function from 5 to up to 20 engineers; owned the technology budget and vendor relationships
Migrated the studio's delivery platform to AWS serverless architecture (Lambda, S3, Elastic Beanstalk), materially reducing operational cost on large-scale marketing campaigns
Shipped computer-vision features using AWS Rekognition as a client offering — early hands-on ML/AI product work predating the LLM era

AWS Serverless S3 Elastic Beanstalk Rekognition React Node.js Vue.js Nuxt.js

Built award-winning WebGL experiences for major film releases including Mission Impossible: Fallout, Terminator Genisys, The Equalizer 2, and Good Omens — Lovie Bronze winner, Webby Honoree
Mentored junior developers into senior roles; led technical prototypes that won new client business
Built full-stack applications for high-traffic marketing campaigns, including a Visit Britain Tumblr campaign that won Pixel Awards People's Champ

JavaScript WebGL Three.js Pixi.js Node.js PHP MySQL

Developer at glue Isobar (Oct 2012 - Jul 2013)
Mobile Developer at Isobar Mobile (Aug 2011 - Oct 2012)

Delivery & Creative Projects

Key Deliverables

Digital Labels System — 6 custom templates across 6 museum sections with multilingual support (196 labels)
Interactive Gaming — Unity-based and web-based experiences with Phidgets sensor integration
Content Management — Custom Strapi CMS for real-time content updates

Technical Highlights

React + Next.js — Template-driven label system with local caching
Phidgets Integration — Physical controls for digital exploration experiences
Docker + Synology NAS — Robust local deployment with hardware teams

Seamless visitor journey with cached, multilingual content accessible across all exhibits.

React Next.js Unity Phidgets Strapi CMS Docker Synology NAS

Key Deliverables

3D Projection Mapping — Unity-based interactive mosque model with animations and exploration
7-Language Digital Labels — Complete multilingual exhibit experience (86 labels)
Interactive Calligraphy — Cultural heritage digital preservation tool

Technical Highlights

Vue + Nuxt.js — Multi-template system for diverse exhibit types
3D Modeling Integration — Interactive architectural exploration with media content
Hybrid Deployment — Windows interactives + Ubuntu thin clients optimisation

Cultural education platform serving international visitors in their native languages.

Vue.js Nuxt.js Unity Projection Mapping Windows Ubuntu

Key Deliverables

Custom iPad Control App — Complete presentation space orchestration
AR Experience Platform — iPad-controlled augmented reality for data visualisation
Hardware Integration — Lighting, audio, and projection system coordination

Technical Highlights

Vue.js + Vite — Wrapped web app within iOS native container
TouchDesigner Integration — Real-time hardware communication and control
Multi-System Coordination — Projectors, lighting, and audio synchronisation

Governmental data storytelling platform highlighting UAE early childhood development initiatives.

Vue.js Vite iOS Native TouchDesigner AR Hardware Integration

Key Deliverables

Interactive 3D Displays — Dynamic Arabic text rendering in 3D space with user interaction
Sensor-Based Systems — Motion and touch-activated playback experiences
Digital Labels Experience — Museum-wide interactive content system