Juan Pablo Gallegos → Work → H2HI
H2HI
Pulling structured insight out of raw audio — back when AI in a product wasn't the default.
What changed before the long read
The case is about turning raw audio into business-readable output when AI product patterns were still immature.
Context
H2HI listens to recorded business calls and turns them into structured, client-specific insight — the themes, signals, and summaries a team can act on instead of re-listening to hours of audio.
It was a full-stack product with a real job: compress conversations into something a business could actually read.
Problem
Raw audio is the least structured data there is. Getting from a recording to "here is what mattered, for this client" means transcribing reliably and then making sense of messy, domain-specific transcripts.
And this was 2021 — before pulling AI into a product was a well-trodden path. There was no obvious playbook to copy.
Constraints
Speech-to-text of the era was imperfect, especially on domain jargon and crosstalk, which set a hard ceiling on the input.
The output had to be readable by businesspeople, not data scientists — no confidence scores, no model jargon.
A small team with full, end-to-end ownership of the stack.
Role
I was a full-stack engineer across the whole pipeline: the React interface, the Python backend, the speech-to-text integration, and the post-processing that turned model output into something a client could read.
System shape
H2HI is a pipeline that takes audio in one end and hands readable insight out the other:
Audio is transcribed by a speech-to-text service, then cleaned, segmented, and mined for client-specific signal in a post-processing stage. The result is structured insight rendered in a React presentation layer.
Because the front of the pipe is messy, most of the engineering lived downstream — in shaping imperfect transcripts into something coherent.
Key decisions
The transcription was never the answer; it was the input. The product's value lived in the post-processing that made messy transcripts legible.
Generic summaries were noise. Insight had to be framed in each client's own terms to be worth reading.
The presentation layer shows insight, not confidence scores or model internals — businesspeople see conclusions, not plumbing.
Trade-offs
Gained
- Hours of audio compressed into readable insight
- An AI-in-product capability years before it was common
- Full-stack ownership kept the pipeline coherent
- A presentation layer business users could actually act on
Cost
- Imperfect transcription set a ceiling on quality
- Heavy post-processing was needed to compensate
- Bespoke, client-specific extraction was hard to generalize
Outcome
H2HI shipped as a working full-stack product that turned recorded conversations into structured, client-specific insight. In 2021 it was also a bet — that AI belonged inside the product rather than bolted on afterward.