ABS Pipeline Architecture

SoS Scrape Pipeline

live API continuous daily harvest · thick pipeline · 7 stages + loops

CT SoS Socrata API →

STAGE 1

HARVEST

Fetch new filings from data.ct.gov · filter shells, non-domestic, bad names · classify: fresh (0-14d) / recent (15-89d) / historical (90d+) · milestone detection

│

sellsat harvest → ct.db →

STAGE 2

BACKLOG + SCORE

Persist in backlog.json · NAICS tier · recency bonus · email domain tier · PLLC fast track (base 92) · priority = score × readiness_weight

│

STAGE 3

ENRICH (7 phases)

1. Filing-service detect · 2. Domain extract · 3. Website check · 4. Phone scrape · 5. Competitor detect · 6. Brave search · 7. Prospeo contact · 8. Equipment mapping · 9. Agent/Principal from SoS · 10. Filing history

│

Brave + Prospeo →

STAGE 4a

RESEARCH

Website scrape · Google Business via Brave · LLM analysis → structured findings · Pushes to Zoho LLMResearch_Note

STAGE 4b

DRAFT

LLM prompt (DeepSeek V3.1) · 7 validation checks: no brands, no placeholders, no fabricated convos · signal tier ★★★ / ★★ / ★ / ☐

│

STAGE 5

SEQUENCE (N×D Accordion)

N touches × D days · exponential gap curve (~1.5x) · min 7d / max 45d gap · purposes: intro → value → proof → offer → urgency → close · 6 presets: light_45 / standard_60 / deep_90 / extended_120 / urgent_21 / renewal_90

│

STAGE 6 — GATE

APPROVE + REVIEW

Interactive queue: a=approve / b=burn / p=pause / d=delete / r=regen · reviewed_at must be set · gate must be OPEN · filing-service emails blocked · suppression check

│

STAGE 7

PUSH TO ZOHO CRM

2-call split · Call 1: all metadata + Sequence + Touch_1-6 · Call 2: Email_Draft_Body2 (send trigger) · Dedup by email → company · Lead_Source = "CT SoS"

│

ZOHO CRM

Email sent via workflow · replies land in email history · polled back by reply handler

SIDE BRANCHES

LOOP

REPLY HANDLING

Polls Zoho · classifies 5 ways: autoresponder (continue), hard opt-out (burn + suppress), soft opt-out (pause), bounce (flag), engagement (respond)

REGISTRY

SUPPRESSIONS

Email + domain blacklist · checked at 4 gates: harvest, renewals, import, push · auto-populated on opt-out

THROTTLE

SEND THROTTLE

Weekly spread: M25 T25 W20 Th20 F10 · max 15/day · scores sorted descending · renewals unthrottled

INTEL

COMPETITIVE INTEL

Agent displacement rates · principal overlap · address clustering · win rate tracking

EXPORT

OUTSIDE SALES

County-grouped field lists · no digital contact → printable ASCII lists for physical visits

COOLDOWN

COOLING PERIOD

120 days after sequence ends · then eligible for new sequence · state machine: eligible → active → paused → completed → burned

STATE MACHINE

Lead Lifecycle

        harvested → scored → enriched → researched → drafted → sequenced → approved → pushed → sent
          |→ burned → suppressed
          |→ paused → resumed
      

UCC Lien Pipeline

periodic export quarterly batches · thin pipeline · 2 stages + push

← EDA Export (18K rows)

STEP A

FILTER & CURATE

CT only: 11K rows → 1K unique · SMB profile: $5K-$75K/item, <$100K total, professional SICs · Past due: 36+ months since last filing → 55 targets

│

├── SELLSAT MATCH ──→ 36% matched

FORK

NAME MATCH vs ct.db

✓ 19 matched via normalization (strip suffixes, punctuation) · full SoS data available

✗ 34 orphans (64%) · no SoS record · created from UCC data only · <OUTREACH ORPHAN> tag in Zoho Description

│

ARTIFACT

STANDALONE CSV (58 columns)

Single source of truth · no backlog · no state machine · no lifecycle tracking · columns: biz+eq+tm+ct+enr+t1-t6+draft+zoho

│

SCRIPT 1

enrich_emails.py

Hunter.io domain search (primary) · Brave fallback · picks best by confidence + title · writes ct_email, ct_email_source, ct_email_verified

│

← Hunter.io + Brave

SCRIPT 2

batch_draft.py

LLM (DeepSeek V3.1) · 6-touch × 45 days · email touches 1,2,4,6 · phone calls 3,5 · saves after each lead · --resume support · writes draft_subject, draft_body, t1-t6 columns

│

← Featherless DeepSeek

SCRIPT 3

batch_push.py

2-call split · Call 1: metadata + Sequence trace + Touch_1-6 · Call 2: Email_Draft_Body2 trigger · Dedup by email → company · Lead_Source = "UCC Lien Data" · Orphan tag in Description

│

ZOHO CRM

Records tagged "UCC Lien Data" · no reply tracking · no suppression check · no send throttle

✗ ✗ ✗ MISSING FROM UCC PIPELINE

MISSING

✗ REPLY HANDLING

No reply polling · no autoresponder detection · no opt-out processing · no engagement response path · replies disappear into Zoho silently

MISSING

✗ SUPPRESSION REGISTRY

No opt-out check before push · no domain blacklist · could re-contact previously burned leads · no compliance layer

MISSING

✗ APPROVAL GATE

No human review before push · no interactive review queue · no burn/pause/delete · draft goes straight to CRM

MISSING

✗ SEND THROTTLE

No weekly spread · no daily cap · no score-based prioritization · pushes all at once

MISSING

✗ RESEARCH STEP

No website scraping · no Google Business enrichment · drafts from UCC + Brave only · less specific content

MISSING

✗ SIGNAL SCORING

No ★★★ / ★★ / ★ tiering · all 55 leads equal · no priority sorting · no readiness weighting

FLAT FILE

State = CSV Row (no lifecycle)

        enriched → drafted → pushed
          (no pause, no burn, no regen, no review)
      

Comparison — what each pipeline has that the other doesn't

Dimension	SoS Pipeline		UCC Pipeline
Data source	Live CT SoS API (data.ct.gov)continuous, automated	↔	EDA XLSX exportperiodic, manual upload
Trigger signal	New business formed → needs first copier	↔	Lease expired → equipment unmaintained → displacement window
Outreach posture	"We exist, here's who to call"seed-planting, long game	↔	"Your equipment financing is expired"urgent, pain-point driven
Storage	backlog.json + ct.dbstateful, queryable	↔	Standalone CSVstateless, flat file
Contact availability	Principal name from SoS filingnamed decision-maker in every lead	↔	64% "Outreach Orphans"no name, no SoS record	KEY DIFF
Match rate	✓ 100%direct from the registry API	↔	✗ 36%name normalization required, 64% unmatchable	KEY DIFF
Enrichment depth	10 phases: website, phone, competitor, Prospeo, SoS agent, filing history	↔	1 phase: Brave (Hunter credits exhausted)no website, no competitor detect
Research	✓ Website scrape + Google Business + LLM analysis	↔	✗ None
Approval gate	✓ Interactive review queue (a/b/p/d/r/q)	↔	✗ None — push goes direct to CRM
Send throttle	✓ Weekly spread + 15/day cap + score priority	↔	✗ None — all at once
Reply handling	✓ Poll + classify (5-way) + response drafting	↔	✗ None — replies go to Zoho silently
Suppression registry	✓ Checked at 4 gates + auto-populated on opt-out	↔	✗ None
Signal scoring	✓ ★★★ / ★★ / ★ / ☐ tiers + readiness weight	↔	✗ None — all leads equal
State machine	8 states: harvested → scored → enriched → researched → drafted → sequenced → approved → pushed → sent ± burned/paused	↔	3 states: enriched → drafted → pushed
Pipeline duration	Continuous (daily harvest) + 120d cooling period	↔	Quarterly batches (per EDA export cycle)
Code footprint	~8,500 lines17 capabilities + CLI + integrations + state	↔	~800 lines3 scripts, no state, no config
Lead_Source in Zoho	"CT SoS" (cold) / "Renewal"	↔	"UCC Lien Data"