I made my portfolio answer questions in my own voice

Retrieval was the easy part. Making a ₹400/month VPS speak fast enough that it didn't feel broken — that took the actual engineering.

pauseddeploy's paused, not the project

95%LESS LATENCY ON REPEAT QUESTIONS

Semantic cache in Upstash Redis keyed on sentence-transformer embeddings. Ask something close enough to a question I've already answered and you get pre-rendered audio instead of a cold inference run.

~118msCACHED ANSWER

3CONTAINERS, ISOLATED

QdrantVECTOR STORE

1 VPSENTIRE FOOTPRINT

FIG.2 — WHAT HAPPENS WHEN YOU ASK IT SOMETHING

01fastapi

you askclean it up, embed it

02redis

have I heard this?close enough → send the audio

03qdrant

go find contexttop-k chunks from my notes

04llm

answer, groundedno context, no claim

05voxcpm

say it in my voicepre-rendered on a local GPU

api — FastAPI, request path only
inference — isolated so crashes stay contained
sync — pushes embeddings + audio to the VPS

DECISIONS, AND WHAT THEY COST ME

nothing here was free

I CHOSE: cache before scaling anything
BECAUSE: Almost everyone asks the same six questions. Generating those answers twice was just me paying for my own laziness.
IT COST ME: stale answers until a re-ingest clears the keys

I CHOSE: generate voice on my own GPU
BECAUSE: TTS on a cheap VPS was the actual bottleneck. Rendering offline and syncing files kept the box small and the bill smaller.
IT COST ME: genuinely new questions don't get real-time speech

I CHOSE: split inference into its own container
BECAUSE: A model crash shouldn't take my portfolio down during the one hour a recruiter is looking at it.
IT COST ME: more compose config, more things to watch

I CHOSE: Qdrant instead of pgvector
BECAUSE: I wanted filtered hybrid search without contorting a relational schema around vectors.
IT COST ME: another service to run, back up, and forget to monitor