Hunter Borman

I run the AI stack inside a 9-figure DTC wellness brand: the strategy, the vendors, the dev team, and a good share of the production code. This site is the roadmap I work from. Six phases that take a store from messy spreadsheets to AI that runs the place, every step proven in production.

          live right now on a 9-figure store:
          AI bundle builder ↗
          brand concierge ↗
          VIP journeys · nightly
        

[ the operating range ]

Strategy and architecture and production code: the interesting part is running all three at once.

boardroom ◂──────────────────────────────────✶────────────────────────────────── ▸ keyboard
the range this work covers

01 · ROADMAP

Map the plan

Audit the stack, sequence the six phases below.

02 · ALIGN

Align the company

Execs, marketing, CX, and engineering behind one sequence.

03 · VENDORS

Sign the stack

Evaluate, negotiate, and own every AI and data vendor.

04 · TEAM

Lead the builders

Hire and run the dev team; today that's 2 engineers at Cymbiotika.

05 · CODE

Ship it myself

Production systems, hands on keyboard. Every phase below is mine.

[ where it shows up ]

Most brands are still piloting AI. I put it into production — where it moves conversion, retention, and ops hours.

for growth

▲
▲▲▲ △
▲▲▲ Grow revenue

AI bundle pages and an on-brand concierge that turn browsers into subscribers.

conversion vs.
standard pages

Live at Cymbiotika — AI bundle builder + brand chatbot.

for cx & ops

╔══╗
╚══╝╔══╗
╚✓ ╝ Cut cost & risk

Automate the weekly ops grind — exports, segments, “personal” sends — and catch what people miss.

saved across
3 acquisitions

−70% legal spend · hundreds of manual processes removed.

for the exec team

» » » Move faster

Decisions that took weeks — deal rooms, catalogs, vendor stacks — read in an hour, with receipts.

docs/deal, read
in under an hour

98.5% match accuracy · −40% compute.

Working on something similar?

Always happy to trade notes on AI in e-commerce, or ask the assistant about any of it.

Say hello

The AI roadmap [ data → autonomy · six phases ]

Most companies buy AI in random pieces. I sequence it — each phase below unlocks the next, and every one is proven by a system I shipped.

▖▕▔ each phase runs in the tool of its era — spreadsheets → AI

01Consolidate
EXCEL 02Structure
IDE · WAREHOUSE 03Automate
SEGMENT In[04]Augment
NOTEBOOK 05Personalize
SHOPIFY ✦06Converse
AI

              X Microsoft Excel — customers_FINAL_v3.xls
              −□✕
            
FileEditViewInsertFormatToolsDataWindowPHASE 01 — CONSOLIDATE

              ▭
              ⤓
              ⎙
              
              Σ
              A↓
              ▥
              
              fx applied to 0 of 4 files
              100% ▾
            

              A5
              fx
              =COUNTIF(files,"jane*") → 4 silos, no shared key
              scroll ▾ or click a sheet tab
            
New workbook
Open…
Publish to warehouse…
Page setup…
Exit
✓ customers_master.csv published · opening the warehouse IDE…
A
B
C
D
E
F
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
01Kill the spreadsheet silos.
four teams · four files · four versions of the same customer → one source of truth
A1:D3
marketing · web_export.xls · edited Tue
jane d.
jane@gmail.com
signed up 3/12
their “Jane”
email team · mailchimp.csv · edited Mon
Jane Doe
jane@gmail.com
opens: high
their “Jane”

                      ✕ no shared key — nothing joins these four
                    
retail · pos_2019.xls · stale 14 mo
J. DOE
card #C-1042
in-store ×2
their “Jane”
app · app_users.db · live
jdoe_92
jane@gmail.com
iOS · daily
their “Jane”

                      
                      one customer, spelled four ways
                      — every team trusts a different file, and none of them is whole.
                    
=EXACT(F5:F8) → TRUE
source
record
email / card ← the hidden key
fingerprint
web_export.xls
jane d.
jane@gmail.com
#A4F9
mailchimp.csv
Jane Doe
jane@gmail.com
#A4F9
pos_2019.xls
J. DOE
card #C-1042 → jane@
#A4F9
app_users.db
jdoe_92
jane@gmail.com
#A4F9
=EXACT(fingerprint) → 4 records · 1 person
the key was hiding in plain sight — same email, same card, four owners. Hash it, and the silos confess.
Chart 1 — copies_per_customerembedded
420
before · 4 copiesafter · 1 record
 per-team files one_identity

                        ✓
                        One source of truth — live for every team.
                      
4 sheets → 1 record · dupes: 0marketing · CX · retail · app: same customerhundreds of manual segmentation tasks — gone
At Cymbiotika this became the enterprise identity graph. But a .xls file is no home for the truth — next: give it a real one.

              ◀ ▶
              
              
              
              +
              customers_FINAL_v3.xls · 100%
            

              4 versions of the same customer · no shared key
              Sum=1 customer
            

               warehouse — structure.workspace
              ⌘⇧P postgres connected↗
            

                ▤⌕⚲▦⚙
              
EXPLORER · WAREHOUSE
▸ customers_master.csv✓
source
excel → warehouse ✓
phase 01 handoff

                  customers_master.csv ✕
                  schema.dbml
                  scroll ▾ or click a file
                
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
// PHASE 02 — STRUCTURE · the upgrade from spreadsheets
02 · One schema everything can trust.
// free-text in → canonical rows out. believable data for every system downstream.
// customers_FINAL_v3.xls → a spine, forever
Table customers {
  id           uuid [pk]
  email        text [unique]
  fingerprint  hash [not null] // #A4F9 lives here now
}
Table identities {
  fingerprint  hash [pk]
  sources      text[] // web · email · pos · app
}
Ref: customers.fingerprint > identities.fingerprint
[pk]primary key
[unique]constraint
[not null]constraint
[note: '…']
⧄ customers_master.csv imported — 4 sources merged · phase 01
✓ dbml valid · 2 tables · 1 ref — the spreadsheet finally has a spine.

                      table: inbound ▾
                      WHERE confidence > 0.95
                       live
                      conf: avg 0.98 · floor 0.95
                    
raw_identifier
canonical
conf
resolved_by
"Sod. Hypochlor."
NaOCl
0.99
✦ AI
"citric acid (anh)"
C6H8O7
0.98
rule
"AQUA"
H2O
0.99
rule
"MgCl2·6H2O"
MgCl2
0.97
✦ AI
"misc slurry #7"
—
0.61
held back
(4 rows) · 41 ms · 1 held back by the confidence floor · 4,000+ new identifiers a day — resolved, believably.
1 — ∞
1 — ∞
identities
fingerprintpk
sources
orders
customer_idfk
sku · total
customers
idpk
emailunique
fingerprintfk
events
customer_idfk
quiz · email · app
products
canonicalfk
taxonomy · formula
subscriptions
customer_idfk

                        
                        match accuracy — every row believable
                        [ −40% compute ]
                        [ incomplete 13% → 4.5% ]
                      
every join here used to be a VLOOKUP between files · in production at Smarter Sorting — the clean layer every downstream system trusted.
TERMINAL · psqlwarehouse · prod
$warehouse import customers_master.csv ······ ✓ 4 sources merged
$segment connect --source warehouse
✓ source connected — opening Journeys…
⎇ main ✓ · dbml valid · 0 problemsLn 12, Col 1 · UTF-8matched 98.5% · −40% compute · Smarter Sorting → prod

               Segment · cymbiotika-prod
              env: prod ▾↗
            
ConnectionsSources warehouse ✓DestinationsJourneys vip-conciergewinback nextquiz-follow-up nextFunctionsAudiencesworkspace
cymbiotika-prod
Journeys /vip-conciergev3 · liveruns nightly · 07:00
                    
                    
                    
                  
 live · scroll ▾
PHASE 03 — AUTOMATE · JOURNEY CANVAS
Automate the busywork.

                    ▸
                    ✥
                    ▭
                  
−  100%  +

                      runbook: vip_emails.docx · every single week
                      owner: whoever remembers
                    
☐Export the VIP list to CSVmanual
☐Cross-check orders across 3 tabsmanual · error-prone
☐Write 300+ “personal” emails from a templateall sound the same
☐Mail-merge, send, hope nothing breaksno undo
…one of hundreds of manual processes across the company. The list never gets shorter by itself.
drawn once on the canvas — then it runs forever:
⚡ order placed
trigger · storefront event
identity enrich
reads the warehouse · 01–02
✦ AI draft
orders · reviews · quizzes
nightly · 07:00
👤 human review
one click: approve
send
312 unique emails
node: ✦ ai_draft
inputs: orders · reviews · quizzes
style: plain-text · no template
grounded by: identity graph (01)
Built on the identity graph from phase 01 · drafts from each VIP's full history — no templates anywhere.

                        
                        personalized notes, written from real history
                        0
                        templates
                      
RUNBOOK RETIRED− export CSV · cross-check tabs · mail-merge by hand · hope
+ the journey runs itself, nightly — a human just clicks approve

                        runs nightly · 07:00
                        human-in-the-loop · one click
                        winback & quiz-follow-up: next in line
                      
Live at Cymbiotika — one of the automated CX workflows that replaced hundreds of manual processes.
✓ busywork automated — next, the hard calls ▸ opening dd_engine.ipynb
● journey: live · human-in-the-loop · proof: vip-concierge312 delivered · 0 templates

               dd_engine.ipynb — Last Checkpoint: the night before close
              Python 3  ↗
            
▶ Run■Code ▾
                
                
                
              
kernel: idle
PHASE 04 — AUGMENT · NOTEBOOK
# Put AI on the hard calls.
Every document in the deal room, read in minutes — the judgment work that used to take a legal team weeks.
In [1]:docs = load("dealroom/") # 5,000+ files · the whole room
Out[1]:indexing dealroom/100%
contracts
legal
finance
HR
vendors
5,000+ files loaded · too many for any human to read before close — so nobody ever did.
Out[2]:HIGHuncapped indemnitycontracts/msa_acme · §9.2
HIGHchange-of-control triggerlegal/loan_covenants · §4.1
MEDauto-renew · 5-yr lock-invendors/dc_lease · §12
+ 72 more, ranked by severity…
75+ critical liabilities surfaced before close — every flag cites its doc & clause.
In [3]:report = consolidate(risks) # the one the lawyers actually read
Out[3]:saved across 3 acquisitions · 5,000+ docs/deal · < 1 hr
false negatives−90%
legal spend−70%

                        ▸ dd_risk_report.pdf — ranked & cited ⤓
                        75+ liabilities pre-close
                      
distributed pipeline · o1-mini over sharded docs · Trilogy
✓ judgment, augmented — now make it sell ▸ opening the storefront…
Python 3 · idle · proof: dd-engine · Trilogy$400K saved · < 1 hr per deal

              S Cymbiotika · Online Store
               published ↗
            
‹ ›cymbiotika.com/pages/create-your-bundle
                
                
                
              
PHASE 05 — PERSONALIZE · ONLINE STORE
Turn intelligence into revenue.
The storefront now reads everything phases 01–04 built — and rebuilds itself for the visitor in front of it.
JDreturning visitor · Jane Doe — #C-1042
the same customer phase 01 unified — now recognized in one hit

                    quiz: stress & sleep
                    bought Magnesium ×3
                    email: high opens
                    app: daily
                  
profile assembled live from the warehouse (01–02) · kept fresh by journeys (03)
✦ composing a bundle for #C-1042…
Magnesium L‑Threonate
✦ AI pick$48
Liposomal Vitamin C
✦ AI pick$62
Omega‑3 DHA
✦ AI pick$54
picked from her quiz, her orders, her reviews — not a bestseller list.
3×standard page1.0×
AI bundle3.0×
conversion vs. standard landing pages — the first phase the business feels in its P&L.
see it live on cymbiotika.com ↗
✓ every system online — launching the assistant…
● online store · published · proof: bundle-builder3× conversion · Cymbiotika

PHASE 06 — CONVERSE · THE DESTINATION
                    
                    
                    
                  
✦ AI-NATIVE↗
06AI becomes the interface.
Every phase you just scrolled — one conversation on top of it. An autonomous, brand-aligned assistant your customers simply talk to.
CONNECTED SYSTEMS
sheets → warehouse01
schema & matching02
journeys · CX03
dd-engine · docs04
storefront · bundles05
orders · subscriptionslive
every connector is a phase you just scrolled.
✦ brand-assistant · grounded in phases 01–05⌘K
online — all connectors green
▍ it can see everything we built — ask it anything
Where's my order — and can I pause next month?
✦ Shipped Tuesday, arriving Thursday — and done, next month is paused. Want me to bump your Magnesium refill instead?

                          ✓ subscription paused
                          undo
                        
answered and acted · streamed by AI · grounded in CX systems · 0.8s

                          01 sheets ✓→
                          02 schema ✓→
                          03 journeys ✓→
                          04 notebook ✓→
                          05 store ✓⇒
                          ✦ one interface
                        
The whole roadmap — one conversation on top.
Messy sheets became a warehouse, a schema, journeys, judgment, revenue — and now a single assistant that can see and operate all of it.

                          
                          try it on cymbiotika.com ↗
                        
 message brand-assistant…▸

                proof · brand-chatbot · Cymbiotika · subscriptions, orders & product guidance
                scroll ▾ or click a stage
              

Hunter Borman

Hunter Borman

Chat with my work