AI systems architect · research → production

Bishal Upadhyaya.

I build AI systems end-to-end —
from peer-reviewed research to products people use.
Currently: Vibeset.

Available for work

Vibeset — my company

One catalog, three products.

I co-own and build Vibeset — AI music tooling that finds the right track, fits it to picture, and proves what it is. Everything below is real: two of the three panels are wired to production systems.

Find · Curation

AI setlist generation over a licensed, deeply-tagged catalog

The engine that started Vibeset. Describe a vibe — artists, genres, moods, BPM — and it assembles DJ-quality setlists: tempo-matched, harmonically compatible, energy-arc aware. Behind it sits a Postgres + pgvector catalog built by our own scraping and enrichment pipeline, searched three ways (SQL, hybrid, semantic embeddings) and finished by an LLM ensemble.

search 1–150 ms · modes sql · hybrid · semantic · sequencing bpm + key + energy arc

Open the DJ app ↳ More on Curation

live demo — searches the production catalog

Fit · Cue

Music perfectly synced to picture

Upload a cut, and Cue reads it — pacing, mood, moments — then matches licensed music to it, sync points included. A FastAPI + Lambda backend runs the video and audio analysis (OpenCV, librosa, multi-LLM); a Next.js app wraps it for creators. Live and in active development.

status live · free for creators · analysis video + audio + LLM

Open Cue ↳ More on Cue

the shipped app — the link is the demo

Cue sign-in screen: “Music perfectly synced to picture.” Cue's curation setup with a video analyzed and ready

Prove · Choon

Audio fingerprinting and provenance, built for repeated scanning

A hybrid identifier: Shazam-style spectral landmarks for the fast path, a 27.7M-parameter Conformer embedding model (FAISS + temporal alignment) when the audio has been mangled — plus audio watermarking with C2PA signed manifests for provenance. Benchmarked at 66,000-track scale for a major music label evaluation.

recall@1 @ 66k tracks 76.9% · params 27.7M (−42%) · core conditions 93.8%

Run the real stress test ↳ Read the 66k-track case note

in-browser workbench — mangle it, then identify it

▶ play the clip, mangle it, then identify it

illustrative — the real matcher runs on GCP

Research — where the signal started

Neurons first, models second.

Before I built AI systems I studied the original ones — electrical synapses in living circuits. Four peer-reviewed papers, from C. elegans connectomes to making deep networks radically smaller.

  1. 2023 A Generalization of Continuous Relaxation in Structured Pruning Nvidia, Thermo Fisher Scientific (arXiv)

    Structured pruning asserts that while large networks enable us to find solutions to complex computer vision problems, a smaller, computationally efficient sub-network can be extracted. We propose a generalization of continuous relaxation in structured pruning to effectively identify these optimal sub-networks, ensuring performance on resource-constrained hardware.

    with Brad Larson, Luke McDermott, Siddha Ganju

    Applied in production: Choon holds 76.9% recall at 66,000-track scale with a 27.7M-parameter network — this paper is why I believed capacity was misplaced, not missing.

  2. 2021 Circumventing neural damage in a C. elegans chemosensory circuit using genetically engineered synapses Cell Systems (University of Washington, Fred Hutch)

    Investigating neural circuit plasticity, this study demonstrates how genetically engineered electrical synapses can reroute information flow to circumvent damaged neurons. By expressing connexin gap junctions, we successfully restored chemotaxis behavior in C. elegans after critical interneuron loss, highlighting the potential of artificial synapses in neural repair.

    with Dr. Sreekanth Chalasani, Research Team

  3. 2020 A Comparison of FDG and Amyloid PET for the Deep Learning Prediction of Alzheimer’s Disease University of California, San Francisco

    This research compares the efficacy of FDG PET (metabolic) and Amyloid PET (pathological) biomarkers in training Deep Learning models for Alzheimer's prediction. The study focuses on applying these models to data from underprivileged communities, evaluating accessibility versus predictive power for early diagnosis.

    with Research Group

  4. 2019 INX-18 and INX-19 play distinct roles in electrical synapses that modulate aversive behavior PLoS Genetics (University of Washington, Fred Hutch)

    Electrical synapses are critical for fast neural communication. This paper elucidates the distinct functional roles of innexin proteins INX-18 and INX-19 in forming gap junctions that modulate aversive behavioral responses in Caenorhabditis elegans, contributing to our understanding of the genetic basis of behavior.

    with Lab Colleagues

Across domains

Voice → structured identity

Golo

Listens to someone talk for two minutes and returns a structured psychological profile — Big Five scores, separate public and private personas — with schema-enforced outputs across multiple LLM providers.

speech → insight · structured outputs · multi-model routing

Source ↳

LLMs under risk discipline

KTM Capital

Reads a week of financial headlines, scores sentiment with LLMs, and paper-trades on it — inside hard stop-loss and position caps. The interesting part isn’t the alpha; it’s trusting a model with money and building the guardrails that earn it.

evals in the loop · hard risk limits · scheduled autonomy

Source ↳

Infra that survives audits

Production discipline

Terraform-managed AWS and GCP, SOC2-minded audit logging, and documented cost-governance sweeps where every deletion was verified recoverable first. The unglamorous half of AI systems, done properly.

terraform · aws + gcp · least-privilege by default · cost governance

Work with me

Bring me the hard part.

Consultation · Free

A quick call to explore fit. If I can't help, I'll say so and point you somewhere better.

Book a chat ↓

Retainer · Monthly most chosen

Ongoing AI advisory — architecture reviews, prototype builds, an extension of your team.

Get started ↓

Project · Custom

Full-scope build, concept to deployment. I own the AI system you need to ship.

Let's scope it ↓

Forms too slow? Ask Balgo upstairs, or go direct: