Get In Touch

AI Product · Case Study

Tandem

Clinically-grounded AI for the parents of twins.

Most baby apps assume one child. Tandem is built from the ground up for two — coordinating both babies into a single sleep schedule and answering parents' questions from a base of published pediatric research, not guesswork.

2.5% hallucination rate, down from 14% — every answer grounded in clinical sources
babies, one coordinated schedule — a problem no major app solves
~217K US households raising twins aged 0–3 — an acutely underserved niche

Built by a market researcher and AI engineer to solve a problem in his own home — as the father of twins himself.

See It in Action

A short walkthrough of Tandem building a coordinated schedule for two babies and answering a parent's question with cited research — the live product, not a mockup.

Product walkthrough — video coming soon

What Tandem Does

Parenting twins is not parenting one baby twice. Two wake windows drift out of sync, two developmental clocks tick at different rates, and a tired parent has to reconcile them in real time. Tandem turns that coordination problem into software.

Coordinated sleep scheduling

A joint scheduler optimizes both children's wake windows into one household schedule — naps, bridge naps, and bedtime — so parents aren't whipsawed between two conflicting timetables.

Answers grounded in research

A chat assistant answers questions from a curated corpus of pediatric sleep guidelines and clinical literature — using retrieval-augmented generation so responses stay tied to real sources instead of generic AI output.

Adaptive, per-child predictions

The model learns each child's evolving rhythm and adjusts to developmental phase — and, for preterm twins, to corrected age. One twin's data never contaminates the other's.

Effortless tracking

Sleep, feeds, and diapers log in a tap and roll up into trends — the data that powers the predictions, captured without adding work to an already exhausting day.

Engineered to Be Trusted

Parents act on this advice at 3 a.m. A hallucination here isn't a typo — it's a safety risk. So Tandem is built the way a researcher builds an experiment: to be measured, grounded, and verifiable.

Grounding

Every answer is retrieved from and constrained to a clinical evidence base. Measured hallucination rate fell from 14% to 2.5%.

Evaluation

A three-tier eval harness — deterministic metrics, claim-level analysis, and human review — scores quality on every change, so improvements are proven, not assumed.

Twin isolation

A hard architectural rule guarantees one child's data can never leak into the other's predictions or answers — the integrity guarantee a twins product lives or dies on.

Safety first

Medical-keyword questions bypass the usual time-weighting and route conservatively, so urgent topics never get diluted by routine logging noise.

Production stack · Python · FastAPI · React PWA · Retrieval-augmented generation · vector search + reranking · built with modern AI tooling.

Market Opportunity

Tandem starts where the pain is most acute — parents of twins — a focused, well-bounded niche with no incumbent built for it. The sizing below is deliberately conservative and US-only: every population figure is drawn from CDC and peer-reviewed sources, and the revenue model uses a modest blended subscription assumption. The broader infant market is the expansion story, not the headline.

TAM $49M / yr SAM $31M / yr SOM $1.6M ARR
US twins only · modeled at a conservative $150/yr blended ARPU
LayerWho it countsMathSize
TAMTotal addressable All US households raising twins, ages 0–5 327,000 × $150 ~$49M/yr
SAMServiceable available US twin households in the core 0–3 sleep window, smartphone-owning 208,000 × $150 ~$31M/yr
SOMServiceable obtainable Realistic 3–5 year capture, ~5% of SAM (range 3–7%) ~10,400 × $150 ~$1.6MARR

SOM range: $0.9M–$2.2M ARR at 3–7% penetration. For reference, the leading generalist baby-sleep app (Huckleberry, single-child) is estimated at ~$7M ARR — so a focused twins product reaching one to two million in recurring revenue is a conservative, well-supported target, not a stretch.

The competitive gap

No incumbent is built for two-baby households. Huckleberry, Napper, and Nanit are all architected around a single child — "add a second baby" means toggling between two profiles with no coordinated wake-window math and no joint bedtime. The only dedicated twins app on the market is a community log-book with no sleep intelligence at all. The offline alternative — a human sleep consultant experienced with multiples — runs $795–$1,915 for a single engagement, and many charge twin families a premium on top. Tandem sits in the open space between a generic single-baby app and a one-time four-figure consultation.

Business model

A premium consumer subscription. The market model above assumes a blended $150/yr ARPU — positioned between leading single-baby sleep apps (~$120/yr) and the one-time cost of a human consultant, and far below it. Twin families are a documented high-spend, high-intent cohort, which supports both a premium price band and stronger willingness to pay than the generalist market.

The expansion path

Twins are the beachhead, not the ceiling. The coordination engine generalizes to any number of children, and the clinical foundation applies to every infant — so the same product extends naturally into the far larger single-baby market once the twins niche is owned. That broader market is upside; it is deliberately not counted in the figures above.

Methodology & assumptions

  • Population. CDC / NCHS 2023 final natality data — ~55,000 new US twin deliveries per year. Living stock = annual cohort × years in the age band (0–5 ≈ six cohorts).
  • Age bands. TAM spans 0–5; SAM narrows to 0–3, the intensive nap-coordination window where Tandem's core value concentrates.
  • Reach. 96% smartphone ownership among parents aged 30–49 (Pew, 2025) — conservative here, since twin parents skew older and higher-income (ART-driven).
  • ARPU. $150/yr blended — between Huckleberry (~$120/yr) and the $795–$1,915 one-time cost of a twin sleep consultant.
  • Penetration. SOM models ~5% of SAM captured over 3–5 years.

Population counts are sourced facts; household-stock, ARPU, and penetration are clearly-labeled estimates. Full citations in the Sources section below.

Brand Identity

"We know your babies — not just babies." Tandem is the only system built for the chaos of raising two, not a single-baby app with a twin toggle bolted on. The brand meets parents in the mess and walks them to clarity — grounded in evidence, and never precious about it.

tandem.

"The Clasp" — two crescents, backs together, meeting at a single gold point. The dot is the bond between siblings: "I've got your back."

Palette

The two children's colors sit a deliberate 50° apart in hue — so a sleep-deprived parent tells them apart by color, not just position. Gold carries the palette's entire warmth: one warm note against two cool ones.

Type

DM Sans — one humanist family at multiple weights, lowercase throughout. The wordmark closes on a gold period. The brand doesn't shout.

Voice

"The friend with twins, tattoos, and a stack of research papers — the one you call at 3 a.m." Direct, honest, warm-not-soft: irreverent enough to swear in marketing, disciplined enough never to swear in clinical guidance.

Refuses

Not a medical device. Not a generic tracker with a twin toggle. Not soft or precious.

built for two shows its work talks like a person

Status & Roadmap

Today

A working production system — clinical corpus, joint scheduler, evaluation harness, and a React app — in daily use in its creator's own home.

Next

Hardening for multiple households and preparing the twins experience for a broader release.

Then

Extending the coordination engine and clinical foundation from twins into the far larger single-baby market.

Skills in Evidence

Tandem isn't a slide — it's a shipped system. Here's the research, analytics, and product toolkit it puts to work, mapped to exactly where each one shows up.

Quantitative Analysis

  • Python — a production Python system: FastAPI, async retrieval and scheduling pipelines.
  • SQL · data extraction — event data modeled and queried in SQL (SQLite / Postgres), feeding daily and weekly analytics rollups.
  • Experimental design · A/B testing — competing prediction strategies are A/B-scored offline via shadow methods and replay comparison before any one ships.
  • Machine learning — vector embeddings plus cross-encoder reranking for retrieval, and adaptive per-child forecasting (EMA with a Holt trend term).
  • Hypothesis testing — every change is validated against statistical quality gates: claim precision / recall and retrieval-drift KL divergence.

Market Research

  • Market sizing · TAM/SAM/SOM — the funnel on this page, built bottom-up from CDC natality data and peer-reviewed twinning research.
  • Competitive analysis — 10+ products and human sleep consultants benchmarked on price and capability.
  • Segmentation & personas — segmented by age band around a sharply defined, underserved twin-parent persona.
  • Voice of customer — requirements drawn from the target user's lived pain and validated in daily, in-home use.
  • Thematic analysis — the clinical knowledge base is organized by LLM-assisted topic clustering across ~47 themes.

Marketing Strategy

  • Positioning — Tandem is positioned as the twins-first wedge between a generic single-baby app and a four-figure human consultant.
  • Differentiation — a defensible gap thesis: every incumbent assumes one child; Tandem owns the two-baby coordination problem.
  • Go-to-market — a beachhead plan: win the acute twins niche first, then extend the same engine into the far larger single-baby market.

AI & Product

  • AI strategy — an evaluation-first approach: ship only what a three-tier eval harness proves improves quality.
  • LLMs & generative AI — retrieval-augmented generation over a clinical corpus, synthesis constrained to retrieved evidence — hallucinations cut from 14% to 2.5%.
  • AI agent development — a production agent: multi-step decomposition, parallel sub-question execution, ReAct-style verification, and safety routing.
  • Product-market fit — acute, underserved pain matched to documented willingness-to-pay versus incumbents and consultants.

Tandem exercises the production, AI, and market-analysis side of the toolkit. The statistical-research side — survey design & programming (Qualtrics), in-depth qualitative interviewing, regression / GLM, causal inference, and conjoint / MaxDiff in SPSS and R — is documented across 10+ peer-reviewed publications and seven years of applied SMB and nonprofit studies. See the research →

Sources

Market figures are built from primary and peer-reviewed sources. The population counts are hard demographic facts; the revenue assumptions are stated and conservative.

Derived figures (household stock, ARPU, obtainable share) are the author's own estimates, clearly labeled as such, built on the sourced inputs above.

Want the full story?

Happy to walk through the product, the architecture, or the underlying research.