AI that runs the business while no one's watching it.
I build production AI systems, sales research bots, CRM enrichment pipelines, RAG search over your own team's data for founders and ops teams done doing repetitive work by hand. Built on n8n, Make, GPT-4o, and GCP, with validation baked in so nothing breaks silently.
Worked With
About
Most operations teams are running on a patchwork: a CRM nobody trusts, research done manually before every call, follow-ups that fall through when someone's busy. It's not that the work is hard, it's that it's repetitive, and repetitive work is where good people burn out and good data goes stale.
I builds the systems that take that work off someone's desk entirely: AI that researches a lead before the call happens, scores how the call went after, and keeps a searchable memory of every conversation so nothing gets re-explained twice. Production-grade, not proof-of-concept validation layers included, so the system doesn't quietly start failing the week you stop watching it.
I built this kind of system in healthcare robotics,Mass Media, Public Sector, enterprise sales operations, Contractual Projects and freelance client work since 2024.
Capabilities
Systems Built
Sales Intelligence System
Reps walked into meetings without full context on the person across the table, and most of what happened in the call was forgotten by the next one.
A four-part n8n system. Before every meeting it pulls email history, a scraped LinkedIn summary, and live company research into one briefing with talking points, red flags, and discovery questions. After the call, it scores performance across four dimensions, drafts (but never auto-sends) a follow-up email, and embeds the transcript so every past conversation becomes searchable. A Telegram bot answers plain-English questions about any past meeting, cited to the exact call and date.
Every meeting is now researched before it happens and searchable after no re-explaining a client's history from memory.
CRM Enrichment & Multi-Platform Content Engine
HubSpot contacts arrived incomplete and inconsistent, with no reliable way to enrich them at scale.
An AI research → normalization → CRM update pipeline with fuzzy mapping and enforced JSON validation, paired with a Make-based content engine that routes, detects language and generates platform-specific content automatically. plus Gemini-powered agents handling customer support and lead qualification.
GarudCV — Medicine Detection
Blocked medicine slots on a packaging line went undetected until downstream.
Real-time computer-vision QA that halts the line via HMI the moment a blockage appears. Deployed on Raspberry Pi and Odroid.
AI-Powered Pick & Place Arm
Robotic arms needed precise, real-time coordinates for pharmaceutical-grade handling.
A coordinate-mapping model driving accurate, real-time pick-and-place in production.
Experience
- Shipped a four-part sales intelligence system that eliminated manual pre- and post-call research for a client, cutting related admin work by up to 80%.
- Integrated LLMs and AI agents into CRMs, messaging APIs, and data enrichment services across multiple client engagements.
- Built and shipped agentic AI systems for customer support, lead qualification, and content generation, cutting manual workload up to 60% via n8n/Python automation.
- Implemented fuzzy mapping, validation layers, and structured JSON enforcement so AI outputs stayed reliable in production.
- Shipped two production computer-vision systems, real-time defect detection and robotic pick-and-place, to edge hardware in a live pharmaceutical packaging environment.
How I Work
Discovery
A short call to find the highest-leverage manual process worth automating first.
Map the workflow
A walkthrough of exactly how the process runs today: tools, handoffs, edge cases.
Prototype
A working version built before full build-out, so we validate the approach early.
Build
Production build with validation layers and error handling from day one.
Deploy
Shipped into the tools you already run on, no new system to learn.
Refine
A tightening pass once real usage surfaces the edge cases a spec can't predict.