Production systems solving real business problems. Each project focuses on practical outcomes — time saved, costs reduced, decisions improved.
266 customer calls analysed without the audio leaving the building
Local transcription and analysis pipeline for voice data that cannot be sent to a third party. faster-whisper large-v3 on owned GPU hardware, deterministic rules separated from model judgement, and PII redaction built into the pipeline rather than bolted on afterwards.
✓ 266/266 transcribed, zero failures
✓ 20.7 hours of audio, zero network egress
✓ Overturned the client's own working hypothesis
faster-whisper large-v3CUDAPythonLocal LLMSQLitePII redaction
Proving generative output actually improved — at volume, without bias
A blind evaluation harness for generative AI: the judge never sees the source, positive controls verify it can tell good from bad, and the load-bearing output is a retelling rather than a score. It has contradicted me repeatedly, which is the only real evidence that it works.
✓ Comprehension 3.3 → 7.8 on a blind judge
✓ Judge self-agreement r = 0.78
✓ 56 cycles documented, including every dead end
Local LLMsPythonBlind evaluationPositive controlsInter-rater reliability
Saving teams 10+ hours/week on meeting follow-ups
Automated pipeline that polls Zoom for new recordings, downloads transcripts, extracts action items and decisions using Gemini AI, and posts structured summaries to Slack within 15 minutes of a call ending.
✓ 30+ client meetings processed weekly
✓ 15-minute turnaround (was 2-day delay)
✓ 23 streaming platforms for distribution
Node.jsGemini 2.0 FlashZoom S2S APISlack APISQLiteVTT Parsing
Serving 30+ custom AI assistants from a single runtime
White-label platform for hosting custom AI agents. Each client gets their own configured assistant with custom system prompts, knowledge bases, and conversation history. Configuration-driven architecture — no code changes needed per client.
✓ 30+ active client configurations
✓ Streaming responses with structured outputs
✓ RAG knowledge base integration
Node.jsTypeScriptOpenAI APIMySQLREST API
20+ agents running unattended, surviving reboots and crashes
Framework that coordinates AI agents running on different physical machines (controller + 2 workers). Phase-based project execution with dependency graphs, error recovery, state persistence, and cross-machine messaging via SSH and SQLite queues.
✓ 3-machine orchestration (controller + 2 workers)
✓ 39 automated tests passing
✓ Full error recovery and state persistence
Node.jsBashSSHSQLitePM2DAG Scheduling
Protecting ad spend across 30+ client campaigns
Real-time monitoring system that detects budget anomalies, lead count drops, and performance shifts across Google Ads campaigns. Alerts team via Slack with confidence tiers and time-based escalation.
✓ 30+ campaigns monitored daily
✓ Anomaly detection with confidence tiers
✓ Automated Slack alerts
Node.jsPostgreSQLGoogle Ads APISlack API
SEO Analysis Pipeline
Full-stack SEO intelligence from crawl to report
End-to-end SEO analysis system: crawl websites, extract technical SEO signals, analyze content quality with AI, generate client-ready reports. Processes hundreds of pages per crawl with structured data storage.
✓ 600+ pages analyzed per client
✓ AI-powered content scoring
✓ Automated report generation
Node.jsCrawleeSupabaseGemini AIPostgreSQL
Website Crawler
Production crawler handling sitemaps, redirects, and SEO signals
Self-hosted website crawler built for SEO analysis. Handles sitemap.xml parsing, redirect chain tracking, robots.txt compliance, and outputs structured data for downstream analysis pipelines.
✓ Handles sites with 10,000+ pages
✓ Full redirect chain tracking
✓ Sitemap + robots.txt aware
TypeScriptCrawleeMySQLPlaywright
Slack Mention Tracker
Real-time team communication monitoring with task integration
Monitors Slack channels for mentions, threads, and action items. Syncs with Proofhub for task creation. Multi-user dashboards with per-person views.
✓ Real-time mention tracking
✓ Proofhub task sync
✓ Multi-user dashboards
Node.jsSlack BoltProofhub APIExpress
Freelancer Market Intelligence
Data-driven competitive analysis from 608 profiles
Automated pipeline that scraped and analyzed 608 freelancer profiles from FreelancerMap and Malt. Identified key success factors: GitHub presence (14.5x elite likelihood), bio optimization (11x length difference), and skill specificity patterns.
✓ 608 profiles analyzed (99.7% success rate)
✓ 14.5x elite differentiator identified
✓ Actionable 6-month playbook generated
PlaywrightPostgreSQLNode.jsStatistical Analysis