Applied AI · agents, voice, video, multimodal
I take AI systems from research to production.
Started with images, then 3D, then audio, then video.
Now agents, wired into real products. Try Balgo.
Vibeset
I cofounded Vibeset Technologies. One product watches a video and picks music that fits, one knows a recording again after the internet has mangled it, one finds the right track from a rough description. Point the same three at support calls or security footage and nothing about the engineering changes.
FIT · Cue
Upload a cut. Cue reads its pacing, mood and moments, then finds licensed music that fits, sync points included.
status live · free trial · analysis video + audio + LLM
live
localmadmom grid · 127.66 bpm · v9
0.76
4 of 7 cuts on a beat
Three of six offsets the ranker picked. The cuts land on a beat wherever you start. What changes is which beats, and what the song is doing there.
PROVE · Choon
Two matchers. Spectral landmarks answer clean audio in milliseconds. A small Conformer, with FAISS and temporal alignment, catches everything the internet has chewed on. Watermarking, so a track can carry its own provenance, is in active development.
recall@1 @ 66k tracks 76.9% · params 27.7M (−42%) · core conditions 93.8%
| Model | Size | Recall@1 |
|---|---|---|
| MuQ-Large | ~300M | 88.2 |
| Choon | 27.7M | 85.4 |
| MERT | 95M | 74.3 |
| NAFP | 19.2M | 63.5 |
The learned channel alone, at a tenth of the size of the model ahead of it.
FIND · Curation
The engine that started Vibeset. Describe a vibe and it builds the setlist: tempo-matched, harmonically compatible, shaped like an evening. Underneath sits a Postgres + pgvector catalog and an LLM ensemble to finish.
search 1–150 ms · modes sql · hybrid · semantic · sequencing bpm + key + energy arc
live
Underneath
Ingest and provenance
connectors, crawlers, rights tracked at the source
Curation Cue Choon
Enrich
ETL, LLM labelling, human review in the loop
Curation Cue Choon
Represent
embeddings, fingerprints, learned features
Curation Cue Choon
Train and compress
fine-tuning, pruning, quantisation, latency budgets
Curation Cue Choon
Retrieve
exact, keyword, vector, diversity, rerank
Curation Cue Choon
Serve and prove
inference under a deadline, eval harnesses, cost governance
Curation Cue Choon
| Capability | Curation | Cue | Choon |
|---|---|---|---|
| Ingest and provenance — connectors, crawlers, rights tracked at the source | leans on | uses | leans on |
| Enrich — ETL, LLM labelling, human review in the loop | leans on | leans on | uses |
| Represent — embeddings, fingerprints, learned features | uses | leans on | leans on |
| Train and compress — fine-tuning, pruning, quantisation, latency budgets | does not use | does not use | leans on |
| Retrieve — exact, keyword, vector, diversity, rerank | leans on | leans on | leans on |
| Serve and prove — inference under a deadline, eval harnesses, cost governance | leans on | leans on | leans on |
Research
Before I built AI systems I studied the original ones — electrical signalling between neurons, in living circuits. Then medical imaging. Then how to make a network small enough to run anywhere. The habits are the same in all three: measure honestly, and publish the row that fails.
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.
The capacity was never missing. It was just badly spent.
In production This is the paper that ships. Choon's identifier runs a 27.7M-parameter Conformer, 42% smaller than the model it replaced, and holds 76.9% recall@1 at 66,000-track scale. See Choon ↳
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.
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.
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.
What I want to point this at next is measurement of the harder kind: attention, behaviour, well-being, the texture of someone's actual experience.
Work with me
Send me a product, a workflow, a dataset, or the problem you can't get anyone to take seriously. I'll usually come back with something you can look at. If it's worth continuing, a short paid pilot; if that works, fractional or ongoing. And if I'm not the right fit, I'll say so and point you to someone who is.
Or skip the form.