AVX-CS-001 · Agentic AI · Portfolio Flagship
Autonomous AI SDR Agent
A configurable multi-agent system that discovers, enriches, scores, researches, and writes hyper-personalized outreach for every qualified lead, grounded strictly in verified facts, with zero fabricated data.
agent pipeline stages
Discover, enrich, score, research, and write: one supervisor-coordinated workflow from ICP to a personalized draft.
explainable fit scoring
Every lead is ranked against the ICP with a written reason, not just dumped into a raw list.
fabricated facts, by design
Outreach is grounded strictly in verified research and traced per decision, an anti-hallucination guardrail, not a best-effort promise.
Executive Summary
What we built, and why it matters.
Sales teams and agencies lose hours per batch to manual prospecting: searching for leads, opening every website, reading reviews, judging fit, and hand-writing emails, while generic mass outreach gets ignored. Avrixo built an autonomous AI SDR agent that runs the entire top-of-funnel workflow end to end: discover businesses matching an ideal customer profile, enrich and score each one, research genuine personalization angles, and draft outreach grounded strictly in verified facts.
Delivered as a live web app, a public GitHub repository, and a documented case study, the system is built as a configurable template: swap the ICP and the offer, and it retargets any vertical, from salons to law firms to DTC brands. The demo instance targets dental clinics for an AI-receptionist / missed-call-capture offer, chosen because the data, from websites to reviews to ratings, is entirely public, making the system free to build, demo, and hand to a client pointed at their own market.
The Core Problem
The bottleneck beneath the business goal.
Manual prospecting is slow and inconsistent: finding qualified leads for a niche takes hours per batch, researching each prospect for genuine personalization is the biggest time sink, and the shortcut most teams take, bad targeting plus generic copy, produces low reply rates and wasted spend. Generic mass email gets ignored, and real personalization does not scale with a human doing it lead by lead.
A basic scraper solves half the problem by dumping a list, but a list alone does not drive replies. The system had to reason about each lead individually: score genuine fit against the ICP, find unique angles instead of generic flattery, and never invent a fact about a business it hadn't actually verified. That last constraint is what separates a demo from a tool a sales team can actually trust to send from.
The Technical Solution
Engineered for accuracy, scale, and trust.
A LangGraph supervisor coordinates four specialist agents through a single lead's journey: a Discovery agent finds N businesses matching the configured ICP via search and directory tools; an Enrichment agent pulls each business's website, contact info, rating, reviews, and services; a Scoring step ranks fit 0–100 with a written reason; and a Research agent digs into recent reviews, service gaps, and news to surface genuine personalization angles before a Writer agent drafts the outreach.
Every output is schema-validated with Pydantic before it reaches a user: thin-data or low-confidence leads are flagged rather than silently passed through, and duplicates are removed before they reach scoring. The Writer agent operates under an explicit anti-hallucination guardrail: it may only reference facts the Research agent actually found, producing a tailored first email plus two follow-ups per lead instead of one generic template blasted at everyone.
The system evaluates itself before a batch is trusted: an LLM-as-judge scores personalization quality and fit-score sensibility on a sample set, and Langfuse traces every agent decision per lead, so a fabricated claim or a bad fit score is traceable back to the exact step that produced it, not a black box.
The whole stack runs on free-tier infrastructure by design: Gemini or Groq for reasoning, Tavily or Serper for search, Streamlit for the UI, hosted on Streamlit Community Cloud, so the same architecture behind the dental-clinic demo can be repointed at any vertical by swapping only the ICP and offer configuration, with results exported to Google Sheets or CSV, or pushed to an n8n-friendly webhook.
What's Inside
Every capability, grouped by workflow.
Lead Discovery & Enrichment
- ICP and offer defined in simple, swappable configuration
- Lead discovery via search / directory tools, matched to the ICP
- Per-lead enrichment: website, contact info, rating, reviews, services, socials
- Structured, Pydantic-validated output exported to CSV / Google Sheet
Fit Scoring & Research
- Fit scoring 0–100 with a written, explainable reason per lead
- Deep per-lead research agent surfacing personalization angles (recent reviews, service gaps, news)
- Dedup and low-confidence / thin-data flagging before a lead reaches outreach
Personalized Outreach
- Tailored first email plus two follow-ups per lead
- Anti-hallucination guardrail: outreach only references facts the research agent actually verified
- Human-review-first design: drafts, never auto-sends
Orchestration & Observability
- LangGraph supervisor coordinating discovery → enrichment → research → writer agents
- Langfuse tracing of every agent decision, per lead
- LLM-as-judge evaluation of personalization quality and fit-score sensibility
Delivery & Reuse
- Streamlit input and results UI: enter an ICP, view the scored lead table
- Google Sheet / CSV export plus an n8n-friendly webhook
- Configurable ICP and offer: repoint the same system at any vertical
- Public GitHub repository with a documented case-study README
Architecture Notes
System decisions that made the product viable.
Public-data only, by design: the system uses only publicly available business data, no scraping behind logins and no personal or private data, respecting robots.txt and rate limits with CAN-SPAM / GDPR-aware anti-spam practices built in from the start.
Drafts, not blasts: outreach is generated for human review before sending; the agent assists a sales rep, it does not mass-blast on its own: a deliberate scope boundary, not a missing feature.
Zero fabricated data as the actual success bar: the system only counts as working if a stranger can open the live URL, enter a niche, and get a usable lead sheet with drafts that reference real, specific, trace-verifiable details about each business.
Deliberately bounded scope: no paid data providers (Apollo / ZoomInfo), free public sources only, keeping the entire stack reproducible at zero cost; no CRM billing, multi-tenant auth, or payments; phone/voice outreach is treated as a separate flagship, not scope creep on this one.
Tech Stack Matrix
Infrastructure behind the outcome.
Reasoning & Search
Gemini / Groq (free tier) · Tavily / Serper · LLM-as-Judge Eval
Agents & Data
LangGraph (Multi-Agent Supervisor) · Pydantic v2 · SQLite
Enrichment
httpx · BeautifulSoup / trafilatura
Delivery & Ops
Streamlit · Langfuse · Google Sheets API / CSV · GitHub (Public Repo)
Start a Conversation
Need a system built like this?
Fixed scope, clear milestones, and a team that ships production systems end to end.