AI & Automation — Selected Builds
I build AI systems that take over the parts of a business that run on repetition.
Sales conversations, outbound pipeline, the front desk, compliance — each one designed as a set of narrow, purpose-built agents rather than a single model trying to do everything. Everything below was designed and built solo.
01
Multichannel Sales Agent
Inbound — WhatsApp, Instagram, Facebook, Email, Web Chat
Unifies every inbound channel a business gets leads on into one system, so a lead is handled as a single ongoing conversation instead of five disconnected inboxes.
What it uses
n8nGPT-4o-miniPostgres memoryMeta Graph APIGmailIntent classifierSlack handoff
How it works
Every message passes through a normalizer that identifies the lead and pulls their history, regardless of channel. An intent layer then routes it to the right specialist:
Sales Agent
Qualifies the lead, moves toward a booked call
→
Product Agent
Educates on what we do, no pressure
→
Support Agent
Resolves directly or escalates by severity
What makes it different
- The Sales Agent mimics real human texting behavior — it splits replies into multiple short messages sent in sequence, like a person actually typing.
- A second version connects live to a client's CRM over MCP, discovering the schema at runtime instead of being hardcoded to it.
Fully built, tested across all five channels. Not yet live for a paying client.
Outbound — intent-based lead generation
Finds companies showing real buying signals and turns them into personalized, ready-to-send outreach — no manual research in between.
What it uses
Make.comApifyClaude HaikuClaude SonnetAnymailfinderInstantly
How it works
Starts from an intent signal — companies actively hiring for a specific role, which usually means budget and growth. From there, each stage has one clear owner:
Research Agent
Finds the domain, identifies the decision-maker
→
Qualifying Step
Verifies a real, deliverable email
→
Personalization
Custom opener from a genuine recent win
→
Follow-up
Loaded into a live, sequenced campaign
What makes it different
- Targets real intent signals instead of purchased contact lists.
- Every message references something true and specific about the company, researched automatically.
- Built as a chain of small, single-purpose agents — cheap to run, easy to debug, easy to swap.
Built and tested end to end on live scraped data. Not currently running on a schedule.
Front desk — booking & scheduling
A front-desk agent that actually takes action — books, reschedules, and cancels appointments, instead of just answering questions and handing off to a human.
What it uses
Low-latency LLM setupSupabase vector storeCRM / calendar toolsTimezone-aware scheduling
How it works
Kept deliberately simple: one agent with direct tool access to the CRM and calendar, so it checks availability and executes a booking, reschedule, or cancellation itself, in the same conversation.
Receptionist Agent
Holds the conversation, low latency
→
CRM / Calendar Tools
Books, reschedules, cancels — no handoff
What makes it different
- Low latency by design — built to hold a live, natural conversation without the lag that makes most bots feel robotic.
- Tool-enabled, not just conversational — performs real actions in the CRM rather than describing what could be done.
Fully functional against live booking tools. Currently a demo build, not deployed for an active client.
Compliance — governance layer for AI systems
A compliance and governance layer that sits in front of any AI system — built for the moment a deal stalls because a client's legal or compliance team needs proof the system is safe.
What it uses
Transparent proxyPII-detection modelsImmutable audit logAlembic policy versioningMulti-tenant deletion
How it works
Every request and response passes through the proxy before reaching the client's system. PII is detected and handled before it can leak, every interaction is logged immutably, and policy is enforced automatically per tenant.
Request / Response
Passes through the proxy
→
PII Detection
Handled before it leaks
→
Audit + Policy
Immutable log, per-tenant enforcement
What makes it different
- Covers GDPR, HIPAA, PCI-DSS, and SOC2 out of the box.
- Policy changes are versioned, never silently overwritten.
- Makes AI adoption sellable to risk-conscious or regulated clients, not just early adopters.
Backend hardening and UI spec complete. Not currently deployed as a live service.
05
RankAgent (In Progress)
Content & Ranking — Generative Engine Optimization (GEO) & SEO
An AI-driven SEO & GEO engine built specifically for B2B SaaS companies. It doesn't just write authoritative blogs and articles—it autonomously engages in highly relevant Reddit threads and X discussions to ensure the SaaS product gets cited naturally across the web, training LLMs to rank it higher.
What it uses
Reddit APIX APIPerplexity SonarClaude 3.5Vector DBJSON-LD Schema
How it works
It builds topical authority through two channels: publishing clustered, high-value articles on the SaaS blog, and deploying autonomous agents to drop helpful, contextual citations in active Reddit and X discussions.
Intent Clustering
Groups SaaS keywords & social signals
→
Content Engine
Writes & publishes authoritative blogs
→
Social Listening
Monitors Reddit/X for relevant pain points
→
Autonomous Engagement
Replies with helpful answers & citations
What makes it different
- Focuses on Generative Engine Optimization (GEO) by actively spreading citations across Reddit and niche communities, training Perplexity and ChatGPT to recommend the product.
- Tailored for B2B SaaS — focuses on high-intent buyer pain points rather than generic top-of-funnel traffic.
- Closes the loop between owned content (blogs) and earned content (social citations) autonomously.
In Progress. Core architecture planned, currently building the social ingestion pipeline.