— receipts

What it looks like
when it actually ships.

A selection. We've redacted client names where we had to. Numbers are real, methods are documented.

All LLM apps ML models MVPs Strategy Automation Research
CupidBot
ML models · MVP · consumer AI MVP → prod

CupidBot — AI that gets you dates

Built from scratch with the CupidBot team: the ranking + chat models behind an autonomous dating-app agent that swipes, chats, and books real dates. Trained on ~200k real conversations; tone-conditioned chat (100+ styles), automated follow-ups, calendar-aware scheduling, and a safety layer that mimics human cadence to avoid platform bans. Featured in Vice, NY Post, Fox, Futurism, Yahoo, BFMTV.

how it works · schematic
01
Profile feed
image + bio
02
Swipe model
trained on your taste
03
Match
mutual interest
04
Chat agent
100+ tones · follow-ups
05
Date booked
calendar invite
read the write-upmethods
Thumbprint
Core ML · computer vision · unicorn startup core system

Thumbprint — automatic furniture placement from floor-plan images

Built the core ML system for Thumbprint (US-based unicorn, commercial furniture design): given a furniture plan as an image, the model parses the layout and auto-places real catalog SKUs in 3D — sized, oriented, and snapped to walls. Powers the platform that has now visualized 181M sq ft and regenerated 428K products across 392K spaces.

how it works · schematic
01
Floor plan IMG
user upload
02
Layout parser
walls · zones · scale
03
Placement model
real SKUs · oriented · snapped
04
3D scene
photoreal · priced
read the write-upmethods
Paysera
LLM apps · regulated · fintech in prod

Paysera — GDPR-compliant customer chatbot

Developed a customer-facing LLM assistant for Paysera (EU payments). Designed for GDPR from day one: PII redaction in the retrieval pipeline, EU-region inference, full audit log per message, right-to-erasure flows wired into the vector store. Citation-grounded answers over Paysera's own help-center, with human-handoff on regulated topics.

how it works · schematic
01
User question
EU region · session
02
PII redact
before retrieval
03
Retriever
help-center · citations
04
LLM answer
cited · grounded
05
Audit log
GDPR · right-to-erase
read the write-upmethods
Cordia
AI strategy & advisory · manufacturing memo → board

Cordia Group — AI roadmap memo for a large building manufacturer

A three-week diagnostic across Cordia Group's plants and divisions: 14 stakeholder interviews, a data-estate audit (ERP, MES, CAD/BIM, maintenance logs, and the inevitable Excel layer), and a 15-page board memo. Not a vision deck — a sequenced list of AI bets, each with an impact estimate, a data-readiness score, and an owner. Excerpts below.

start now — Q1

Quote & tender automation first: RFQ parsing + BOM matching on ERP data that was already clean. Defect detection second — line cameras were already installed, labels cheap to collect. Both shippable inside one quarter.

estimated impact

Quoting cycle ~10 days → under 2, so 30–40% more tenders answered by the same team. Scrap on camera-monitored lines down an estimated 8–12%. Payback inside the first year. Estimates, labeled as estimates — not promises.

data ingestion — what to avoid

No central data lake first (the classic 6-month trap). No OCR-ing 20 years of paper archives upfront — cost exceeds value. No training on pre-2022 sensor data: uncalibrated clocks, silent drift. Start with the two systems that already share keys: ERP + MES.

wait / kill

Wait: demand forecasting, until 12 months of unified cross-division sales data exists. Kill: the internal chatbot over undocumented processes — automate the documentation first, then revisit.

how it works · schematic
01
Interviews
14 stakeholders · 3 weeks
02
Data audit
ERP · MES · CAD/BIM
03
Ranked bets
impact × data readiness
04
Board memo
15 pages · sequenced · owned
read the write-upmethods
B2B staffing
AI process automation · B2B services · redacted 4 systems · live

B2B staffing firm — the revenue grind, automated

For a European B2B staffing firm (name under NDA): four automations shipped in one quarter. An outreach agent that researches every prospect before writing a word. A Telegram helper the partners actually use. An automatic review loop. And a SEO & GEO agent that keeps the firm visible in Google — and cited when ChatGPT gets asked who to hire.

lead gen & reach-out

ICP-filtered company list, enriched and researched by an agent before first touch — recent funding, hiring pages, tech stack. Reply rate 0.7% → 4.9% in eight weeks; ~30 qualified calls booked per month, on autopilot.

chatbot-helper in Telegram

Lives where the partners already live. Drafts replies, preps call briefs from CRM history, logs every touch back automatically. Zero new dashboards to learn — adoption was day one.

reviews & AI visibility

Review requests fire on placement success, routed to Google: 4.1★ → 4.8★ on 5× the volume. The GEO agent keeps service pages structured for AI answers — the firm now appears when ChatGPT and Perplexity are asked for staffing partners in its region.

how it works · schematic
01
ICP list
filtered · enriched
02
Research agent
site · news · hiring signals
03
Outreach
personal · multi-channel
04
Telegram helper
briefs · CRM sync
05
Reviews + GEO
4.8★ · cited by ChatGPT
read the write-upmethods

Want the unredacted version?

We'll share named references on a call once we know your project's a fit.

Contact us →