AI Engineer on the GTM team at Fello. I own the systems the revenue motion runs on, end to end: design, build, ship, and the pager when they misbehave. Sales, marketing, and the data layer underneath both.
Most of my work is private company code, so I write each system up in public instead.
| What it is | Why it's worth your five minutes | |
|---|---|---|
| Revenue OS β | A daily work list for sales reps. Each morning one screen says what to do today, in order, and why. | The product bet, a layered architecture with an enforced boundary ratchet, and a data dictionary where every fact carries the query that measured it. |
| Felix Sells β | An AI teammate that runs signal-driven outbound over email and SMS. The outreach is the demo. | The LLM picks intent; deterministic code gates every send. Frozen evals, shadow-to-enforce rollout, and a privacy wall in the memory. |
| GEO Audit Playbook β | How to measure whether AI assistants recommend a business, from my closed side business to agent skills tuned against real websites. | A real failure (0 clients) told straight, a scoring rubric you can read, and the anti-patterns behind each iteration. |
Each has a founder section and an engineer section; the first two open with a demo video.
| ποΈ Autonomous voice agent | Built to cold-call real estate hiring managers over Twilio + ElevenLabs and book its own meetings. Do-not-call, call windows and disclosure rules are gates in code, not prompt instructions. The same gate design now guards the email and SMS outbound in Felix Sells. |
| π A GTM signal no vendor sells | Infers real estate team size from a name and a company. 117 tests, 1,534 req/s at concurrency 100. A later agentic version enriched 1,755 records in 86 minutes: 1,395 high-confidence, 323 medium, 32 with no answer, zero pipeline failures. |
| βοΈ Arthur: a blog agent that corrects itself | Writes and publishes production articles with minimal human intervention. Each draft runs a bounded loop of writer β fact-checker β guardrail β scorer until it clears the bar, and a deterministic link allowlist makes hallucinated URLs impossible to publish. Reviewer feedback then drives a continuous-improvement loop that traces each repeated edit to the one pipeline stage responsible and proposes a better prompt. Now spans blog, email and ads from one campaign brief. I wrote the first generations of the pipeline, from n8n through three production services; the current platform is a shared build with one teammate, who led the feedback layer. |
| π 30-day sales competition platform | Ran a real cohort: automated bracket tournament, about 3,969 scored entries, and an AI cold-call practice coach over realtime voice. |
- Put the invariants in code, never in the prompt. Every dial path goes through one audited function. It checks do-not-call, attempt caps, live-call detection and the contact's local call window before the phone rings, and writes an audit row for every evaluation, operator overrides included. A prompt can't guarantee anything, so I don't ask it to.
- Treat prompts as deployment artifacts. Changes ship only after clearing 59 frozen evaluation cases (abstention on missing facts, objection reframing, timing clamps, hard compliance refusals), plus a live conversation simulation that caught a regression every unit and integration test passed.
- Design for the model being wrong. Sliding-window health checks instead of consecutive-failure counting. Never retry a billing error. Sentinel values instead of
NULL, so batch queues drain. A named failure taxonomy, so "success rate is 40%" becomes a ranked list of what to fix next. - Write the postmortem into the repo. The next engineer inherits what I learned instead of relearning it.
- Work like a forward-deployed engineer. Sit with the people who use the thing; own the outcome, not the ticket.
In voice, the binding constraint is latency: the inbound personalization webhook has a hard ~1.5 s budget with no room for an LLM call, so everything known at pickup must be precomputed.
- Fello Β· AI Applied Engineer Β· Sep 2025 β now. Owns the GTM AI systems above.
- Monarch Labs Β· AI Engineer (freelance) Β· May β Sep 2025. Built a viral-story-to-newsletter pipeline on n8n. The client reported getting 3β4 hours a day back. This is where I hit n8n's limits, and why I now port fragile visual workflows into real services.
- RePut.ai Β· Junior Software Developer Β· Feb β Jun 2025. Hiring-flow automation (Slack, ClickUp, Calendly, DocuSign) and a voice-chat-with-PDFs RAG tool.
- Vellore Institute of Technology Β· B.Tech, Computer Science Β· 2021 β 2025
- Salesforce Certified AI Associate
Looking for someone to own an AI system end to end? Let's talk.