Production AI Systems That Actually Work
I build LLM applications, RAG systems, autonomous agents, and voice AI — engineered with guardrails, structured outputs, and deterministic state machines.
Technical Skills
What I bring as a AI engineer
LLM & Foundation Models
RAG & Vector Systems
Voice AI & Audio
AI Infrastructure
Architecture
How I architect ai systems
Data Ingestion
Parse documents (PDF, HTML, audio) into structured chunks with metadata preservation and semantic boundaries.
Embedding & Indexing
Generate dense vector embeddings via OpenAI/Cohere, store in pgvector with HNSW indexing for sub-100ms retrieval.
Hybrid Retrieval
Combine dense semantic search with BM25 keyword matching, apply reciprocal rank fusion and cross-encoder reranking.
LLM Reasoning
Route through Claude/GPT with deterministic prompt templates, structured JSON output, and citation enforcement.
Guardrails & Output
Validate via Pydantic/Zod schemas, apply hallucination prevention, and stream results with source attribution.
Featured Work
Production projects as a AI engineer
Planfluence
Meetings into Momentum — AI Meeting Intelligence to Jira Backlog & Slack Recaps
Problem
Engineering discussions frequently end with action items and architectural commitments buried in unreviewed recordings or chat threads. Translating spoken consensus into Jira tickets with acceptance criteria wastes hours of developer and manager time.
Solution
Architected a 4-step pipeline: (1) Capture calls via Google Meet Chrome extension without intrusive bots; (2) Transcribe and extract decisions using Whisper and LLMs; (3) Review, approve, and publish tickets directly to Jira Cloud/Data Center and Trello backlogs; (4) Broadcast structured recaps to dedicated project Slack channels.
Architecture
- Chrome Extension (Manifest V3) capturing tab audio locally with zero bots joining user meetings
- Next.js 15 App Router web application with streaming AI summaries and interactive decision editor
- High-speed Whisper transcription & structured JSON extraction for action items and acceptance criteria
- Bi-directional Jira Cloud & Trello API integrations with one-click backlog and sprint publishing
- Slack webhook automation broadcasting project-specific meeting recaps with deep links
- Lemon Squeezy merchant-of-record subscription billing engine
Verified Results
GeniSteps
Hardware-Accurate Step Tracking & Vitality Companion — 100% Local-First
Problem
Mainstream pedometers drain device battery, lock essential 24-hour trends and data exports behind $39–$79/yr recurring subscriptions, and harvest sensitive biometric location and movement data to proprietary cloud servers.
Solution
Built a silicon-native motion engine with a 4th-order DSP bandpass filter (0.8–3.5 Hz) tuned to human gait biomechanics. Features silent lock-screen updates via >95% IPC throttling, natural stride pacing (~320ms), and 100% data sovereignty in on-device SQLite WAL mode.
Architecture
- Silicon coprocessor integration: Apple CoreMotion (CMMotionActivityManager) & Android Sensor Hub
- 4th-Order DSP digital signal processing filter (0.8–3.5 Hz) isolating true strides from vehicle vibration
- Silent lock-screen live activity widgets with >95% IPC throttling for battery preservation
- Local-first SQLite database running in WAL mode with instant 1-tap CSV and JSON spreadsheet export
- Behavioral habit psychology engine: automated bankable streak freezes and rest-day scheduling
- Standardized machine-readable /llms.txt factsheet for generative engine and AI model evaluation
Verified Results
Kookabe AI Suite
Enterprise WhatsApp RAG Chatbot & Automated AI Voice Mail
Problem
Companies received hundreds of repetitive WhatsApp voice notes, PDFs, and product inquiries daily, resulting in 4-8 hour response delays and missed sales leads.
Solution
Built a hybrid RAG system with pgvector and Claude 3.7 that parses customer inquiries, searches internal PDF documentation, and generates accurate answers with human-handoff triggers.
Architecture
- WhatsApp Business API webhooks ingesting incoming text, voice notes, and document files
- Hybrid RAG combining pgvector dense embeddings with BM25 keyword reranking
- FastAPI backend handling asynchronous LLM streaming and conversation state
- Automated audio transcription pipeline extracting lead data directly into CRM webhooks
Verified Results
Experience
Where I've applied these skills
May 2026 — Present
Delivering production AI workflows, full-stack Next.js/NestJS platforms, and cloud DevOps systems for high-growth tech startups and international clients.
- Consulting on AI agent architecture, hybrid RAG pipelines, and conversational voice systems.
- Building scalable REST APIs, microservices, and database schemas with PostgreSQL and Prisma.
- Maintaining a 100% Job Success Score with verified 5-star client contracts.
Open Source
Related open-source work
CodeCraft
Open-source Lovable alternative — full-stack AI website builder that prompts, edits, and deploys production-ready apps directly to any VPS.
GitHub PR Commenter
Automated LLM code reviewer bot for GitHub pull requests with syntax validation, static analysis, and actionable suggestions.
Process
How I work
Understand the Domain
Map your data sources, user queries, and accuracy requirements before writing a single line of code.
Architecture Design
Design the retrieval pipeline, model selection, embedding strategy, and guardrail layers.
Iterative Development
Build in sprints with measurable retrieval accuracy metrics, not just vibes.
Production Hardening
Add caching, rate limiting, fallback models, monitoring, and cost governance before shipping.
FAQ
Frequently asked questions
What makes your RAG systems different from basic implementations?
I build hybrid RAG with dense embeddings + BM25 sparse search, reciprocal rank fusion, and cross-encoder reranking. Every answer includes strict source citations and hallucination prevention guardrails — not just naive vector similarity.
Do you build multi-agent systems?
Yes. I use LangGraph for cyclical state machine architectures with persistent checkpointing, dead-loop prevention, and explicit human-in-the-loop approval gates for critical operations.
Can you integrate AI into an existing product?
Absolutely. I specialize in embedding AI capabilities into production SaaS — streaming completions, token-budgeted prompt caching, semantic caching with Redis, and graceful degradation when model providers experience latency.
What about voice AI and audio processing?
I build voice qualification pipelines — WhatsApp audio transcription via Whisper, intent extraction into structured CRM data, and natural voice synthesis responses with ElevenLabs.
Need a production AI system built?
From hybrid RAG to autonomous agents — I engineer AI that works in production, not just in demos.