LabFlow AI

A week at Universidad de los Andes, Bogotá

The lab notebook that listens, organises and analyses.

STACK
Next.js · TypeScript · Tailwind CSS · TanStack Query · Plotly · Python · FastAPI · SQLModel · Supabase · ElevenLabs · Ollama Cloud · Claude
STATUS
Live · delivered October 2026
LabFlow AI screenshot

OVERVIEW

LabFlow AI is a voice-first lab notebook for chemical engineering students. In a lab practice, measuring and writing compete for the same hands: gloves on, reagents out, values copied into Excel later with mistakes, and missing readings found only at the end. With LabFlow you dictate “trial 3, temperature 65 degrees”, it saves what it is sure of, asks when it isn't, and shows a trials table, a chart and a summary straight away. Built during Misión TEC × Uniandes, a week at Universidad de los Andes in Bogotá for a course in AI and software engineering, and delivered on its last day.

WHAT I FOUND → WHAT I DECIDED

  • A language model can read a number nobody saidThe model only proposes a tool call; code checks that every number appears in the user's own sentence, converts units, applies ranges and decides what is saved
  • Spoken and typed capture would drift apart if each had its own logicOne turn service for both: the ElevenLabs agent only does speech to text and back, with our API as its custom LLM, so the spoken reply is exactly our text
  • Asking to confirm every value slows a lab down, but a wrong value saved silently is worseA new value in an empty cell saves at once with Undo; one that replaces a saved value, falls out of range or was converted waits on a card that asks first
  • A reply had to arrive within 5 seconds of the end of dictationA fast model first (DeepSeek V4.1 Flash on Ollama Cloud), with Claude Haiku as a time-limited fallback, logged every time it's used

WHAT IT DOES

NOTEBOOKS
Create a notebook with its variables, units and expected ranges, or let LabFlow read the lab handout and propose them, quoting only the ranges it states.
DICTATE
Hands-free voice or typing in Spanish, with numbers in words, decimals with “coma” and “grados” read as °C.
CONFIRM
A card shows what you said next to what LabFlow understood, and asks before replacing a saved value.
WARN
Out-of-range values and unit conversions are flagged before anything is saved.
RESULTS
A trials table, an X against Y chart and a per-variable summary, exportable to CSV.
PHONE
A phone layout with a bottom tab bar, and the screen kept awake while a voice conversation is live.
ACCOUNTS
Sign in with Google or email; each account sees only its own notebooks.

HOW IT WAS BUILT

Planned as releases on a living scope — typed capture with confirmation first, then voice, the chart and warnings — with a written spec for each decision before any code. Built by a two-person team: I led text and voice capture and the notebook pages; my teammate built the phone version and accounts. A FastAPI service on Railway with Supabase Postgres sits behind the Next.js app on Vercel, checked by about 490 API tests, 570 web tests and 42 Playwright phone tests on every pull request.

SCREENS