Kevin
Zou

Electromyography

• EMG silent-speech research at Stanford: a wearable that reads muscle activity off the skin and decodes words said silently. Electrodes → signal → decoder, on a 16-channel OpenBCI Cyton+Daisy rig, with an MCP server so an LLM agent can operate the hardware itself and a real-time labeled-recording pipeline behind a from-scratch decoder.

Machine Learning & Data Modeling

• Built the decoder's modeling stack: EMG-specific data augmentation (channel dropout, time-warping, amplitude jitter, electrode-shift simulation) and vector-quantized tokenization (learned codebooks over the raw signal, HuBERT-style) that turns continuous EMG into discrete units a sequence model can decode.
• Held to a strict leak-free protocol: VQ / SSL tokenization only ties the strong raw-feature baseline rather than beating it, so the real ceiling is sensing, not modeling (125 Hz sits below this signal's Nyquist floor), and the next gain comes from the sensor.
• Current bet: noisy-channel decoding with an LLM prior. Partial EMG evidence × P(words | context) → open-vocabulary intent, not a fixed menu of gestures.

Benchmark · Meta emg2qwerty

• Benchmarked the from-scratch decoder against Meta's emg2qwerty, their open surface-EMG decoding baseline, on a strict leak-free, subject-held-out split, with no train/test contamination, so the comparison holds up rather than being cherry-picked.
• It outperforms the baseline, reaching 95% on a working command set.
• The edge is the modeling stack, not scale: EMG-specific augmentation, VQ tokenization, and eval discipline on a narrow, high-value task.

Preference Engine · Attractiveness Scoring

• A matchmaking engine needs a calibrated per-user scalar for physical attractiveness: subjective, fine-grained, and only comparable within a gender×ethnicity bucket (8 supported). Gold labels come from a per-user Elo fit on 6,231 human pairwise comparisons, with 887 men scored at ≥8 duels each.
• Naive scoring fails: the deployed holistic rubric correlates with human Elo at only ρ = 0.315. A multimodal LLM asked for an absolute rating collapses onto a few round numbers.
• The fix is anchored pointwise scoring: show the judge six human-calibrated anchor faces from the same bucket, then ask for one rating on the gold Elo scale. That reaches ρ = 0.62–0.64, beating every compare-then-aggregate scheme (adaptive Bradley–Terry tops out at 0.50) at ~1/17 the model calls.
• Blending in an identity-embedding Bradley–Terry head, trained on raw duels over 512-d ArcFace / AuraFace embeddings, lifts correlation to ρ = 0.698 on a 523-user held-out set: the embeddings carry within-bucket signal the language pathway can't verbalize.
• We hit a hard perceptual resolution ceiling: the judge emits only 112 distinct scores across 538 users, and on near-equal faces it agrees with humans at 53.6% (chance). On a subjective attribute, perception caps the judge, not prompting.
• My part, the face + embedding pipeline underneath: SCRFD / AuraFace detection, owner-isolation blurring so only the right face in a group photo is scored (947 of 1,013 cohort photos retain a usable face), face-gating, and the identity-embedding infra the Bradley–Terry head runs on. Bias handled as a constraint: per-bucket calibration, no cross-bucket comparison, and equity gaps tracked where the data is too sparse to be decision-grade.

research at Ditto

Messaging Infrastructure

• The product is messages in and messages out, so the backend is a messaging system first. One unified send API sits in front of every channel (iMessage, Instagram DMs), so product code calls a single interface and never talks to a network directly.
• Inbound. Platform webhooks hit a normalizer that flattens each provider's payload into one internal event shape, resolves it to a conversation, and hands it to the agent, with idempotency keys so a retried or duplicated webhook never double-fires.
• Outbound routing. A routing layer spreads sends across a pool of senders with per-sender queues, capacity scoring, rate-limit tracking, and health checks, so traffic flows to the healthiest sender and one degraded path is isolated instead of taking the funnel down with it.
• Real endpoints at scale. Delivery runs on a managed fleet of phone numbers / SIMs across Mac-based workers and modems, each with its own session, health, and rate-limit isolation, so the agent sends from real endpoints across iMessage and SMS without any single number becoming a bottleneck or a point of failure.
• Delivery. Retries with backoff, durable session/conversation state, and worker pools; replies stream out as multiple |||-delimited bubbles, first bubble in a few hundred ms, so a thread reads like a person typing, not a wall of text.
• Correctness under concurrency. A wrong conversation-to-identity mapping would route one person's message to another, so identity resolution and state transitions run atomically: a single DB transaction, no unsafe fallbacks.