80 dBA SNR · 132 dB DR · 146 dB AOP. In a pendant.
We're building the smallest portable recording studio anyone's ever worn, and putting AI on top of it.
LaFayette Labs is a one-principal studio. There's no sales deck of imaginary clients here, only the work in flight. Right now that work is Panakoes: a 4-microphone AI pendant engineered to capture audio at studio-condenser SNR. Every claim on this page is traceable to a commit, on a public repo, with a CI badge next to it.
PANAKOES
An AI pendant that hears like a studio.
Panakoes (constructed Greek for "all-hearing") is a wearable AI pendant. On the surface it's positioned as an AI note-taker in the segment served by Plaud Note Pro and Limitless: BLE sync, transcription, summarization, action items. Underneath, the hardware is a studio-grade portable field recorder. Drop the device in a pocket at a concert and it captures every instrument with usable separation. Wear it to a multi-person dinner and every voice across the table is intelligible. Walk through a forest and the ambient soundscape is recorded with a level of detail no other AI wearable on the market can touch.
The marketing message is "AI pendant with great audio." The reality is "the smallest portable recording studio you've ever held, with AI on top."
Studio condenser numbers, in a 35 × 7 mm pendant.
80 dBA SNR. 132 dB dynamic range. 146 dB acoustic overload. Four microphones, TDM8, 24-bit at 96 kHz, in a 35 × 7 mm enclosure on a 600 mAh cell. These aren't headline numbers waiting on a tape-out. They're the validated targets the rest of the architecture is built to keep.
- A1 4-microphone studio-grade capture array
A wide-stereo rectangular array on a 22 × 18 mm polyimide FPC, four channels multiplexed over TDM8. The mics deliver 80 dBA SNR, 132 dB dynamic range, and a 146 dB acoustic overload point. That's studio-condenser performance, picked specifically so a kick drum at three feet and a whisper across a dinner table both land cleanly in the same file. Part selection is the moat. Specifics ship with v1.
- A2 Low-power BLE 5.4 MCU
Generous internal SRAM enables ping-pong DMA buffering that absorbs worst-case microSD garbage-collection stalls without adding a dedicated PSRAM chip on the audio path. The result is one fewer chip in the audio path, a lower BOM, and one less failure mode in the field.
- A3 microSD slot, USB-C, FRAM session metadata
Push-push microSD socket for user-supplied 64 to 256 GB cards (field-mode 24-bit / 96 kHz / 4-channel writes ~4 GB/hr). USB-C for charging, file transfer at USB 2.0 speeds, dev/debug, DFU bootloader. Non-volatile FRAM stores cursor position so the device survives power loss without losing its place mid-recording.
- A4 35 mm × 7 mm pendant form factor, 600 mAh LiPo
Established by feasibility analysis. Smaller doesn't fit the BOM; larger compromises the wearable thesis.
What the four microphones actually do.
Mic specs are necessary, not sufficient. The four-microphone array unlocks a DSP layer that single-mic pendants physically can't run. This is where Panakoes leaves the competitive set behind.
- A5 Hears the speaker, not the room.
The 4-mic array is engineered to electronically steer its pickup toward whoever is talking and reject sound coming from other directions. In a loud restaurant, the person across the table comes through clean while the next booth fades into the background. Single-mic pendants average the whole room into one blurry track, which is why their transcripts fall apart the moment a coffee grinder starts.
- A6 Knows the difference between a meeting and a forest.
Panakoes is designed to ship with distinct on-device profiles for voice, field, and music capture, each with its own gain curve, EQ, and noise handling. Boardroom dialog gets tuned for intelligibility. A hike gets tuned to preserve birdsong and wind without crushing the dynamics. A live show gets the full 132 dB of dynamic range. One device, three honest recordings instead of one compromised one.
- A7 Keeps people in their seats.
Because four microphones see the room geometrically, the device is engineered to preserve who spoke from where rather than mashing every voice into one mono channel. At a six-person dinner, the transcript can attribute lines correctly instead of guessing. Cloud diarization gets dramatically more accurate input to work with, and competing pendants with one or two mics can't reconstruct this signal after the fact.
- A8 Suppresses noise without sanding off your audio.
Most always-on recorders apply blunt noise reduction that smears consonants and kills room tone, which is why their playback sounds like a bad phone call. Panakoes is engineered for context-aware suppression that recognizes the difference between an HVAC drone and a human voice, and only attenuates what doesn't matter. The result is recordings that still sound like the place they were taken in, not like a denoised artifact.
- A9 Strong wind still sounds like wind, not a hurricane.
Wind and handling noise hitting a microphone destroys recordings. Single-mic pendants have no defense, so users learn to point them away from the breeze, which defeats the wearable thesis. The 4-mic array lets the device detect which mics are getting hit by wind or handling and de-weight them in real time, blending in the cleaner signal from the others. A walk in 15 mph wind ends up sounding like a walk in 15 mph wind, not like a hurricane shredding a single capsule.
Open-source, audit-ready, swap-the-model-next-year.
The cloud backend is MIT-licensed and event-driven on AWS. Two transcription paths: async batch for uploaded files and live streaming for the wearable. The transcription model is pluggable, so the platform survives the next state-of-the-art swap.
- B1 Async batch
S3 upload event → Lambda validates and enqueues → AWS Batch on EC2 Spot GPU runs Whisper-large-v3 (fp16) → S3 transcript → DynamoDB stream → Lambda summarizer (Claude) → RDS metadata → notification.
- B2 Live streaming
API Gateway WebSocket → Session Manager Lambda → ECS-managed GPU Spot instance spawned per session → faster-whisper-large with Silero VAD → WebSocket back to client. Sub-second latency target.
- B3 AI summarization, two tiers
Claude Haiku for fast standard summaries on the free tier. Claude Sonnet for deep summaries and action-item extraction on paid tiers. Cost discipline up front; differentiated quality on the upgrade.
- B4 Auth, audit, observability
Better-Auth for identity, RBAC, step-up MFA. DynamoDB application audit log plus CloudTrail for AWS-API events. OpenTelemetry instrumentation with CloudWatch and X-Ray as backends. Logs archive to S3 and stay queryable in Athena.
- B5 Stripe-billed tiers
Free, Pro, and Team tiers wired through Stripe with real coupon, refund, and dunning flows. Production billing is a feature most pre-alpha projects skip; this one ships with it because the operator discipline costs nothing to do early and everything to retrofit later.
Cloud is in heavy active development with a public CI pipeline (pytest, vitest, Terraform, CodeQL, OpenSSF Scorecard). Hardware repo stays private through v1 launch. Production hardware tracks 2026 to 2027 (target). If you need silicon in your hand this quarter, this isn't the engagement. If you want a studio that'll still be shipping in 2027, it is.
If you read sections 02 through 04 and your reaction was "I need this person on a real system," write directly: phil@lafayettelabs.com. First reply within one business day.
Have a system that has to ship, and a deadline that already moved once?
phil@lafayettelabs.comSend a two-paragraph brief: what you're building, what the constraint is. The studio takes 1 to 2 new engagements per quarter alongside Panakoes. AI-leveraged software, hardware and firmware, fractional advisory. First reply within one business day.
Open to principal-level conversations.
phil@lafayettelabs.comSend the role brief in the body. Phil's open to senior-engineering and principal-level conversations alongside studio engagements. The studio model and a senior in-house role aren't mutually exclusive.