Vectr

What it was

Vectr was two products under one mission: cut the busywork that keeps EMTs from focusing on patients.

• Scene Intel. Satellite + street-level imagery, crew notes per address, transport routing with hazard zones, golden hour countdown, offline-first storage, and a landing-zone (LZ) detection system that auto-scored 100×100 ft helicopter sites from satellite imagery.
• AI Scribe. Ambient voice capture on iOS that converted patient encounters into structured NEMSIS ePCR fields, attacking the ~45 min/shift charting tax. All inference on-device (~450 MB total): Parakeet TDT for speech, LFM2-350M for structured extraction. PHI never left the device.

Stack: Svelte 5 + Tauri on the desktop, SwiftUI + CarPlay on iOS, Rust + Actix + PostgreSQL/PostGIS on the server (~35k req/s), DINOv2 trained on DGX Spark for satellite LZ scoring.

Origin

The idea didn't come from a brainstorm. It came from a 90-minute conversation with Kitty-Jo Cox, a 3-year EMT in rural Tennessee. She walked through a typical night call: dead-end road, no streetlights, fog, no number on the house, truck wheels sinking in a wet yard, every minute without compressions dropping survival 7-10%.

She wasn't complaining about navigation. She was describing a visibility problem. EMTs respond to scenes they've never seen, in conditions where their tools (a $100k CAD system, a 4 GB Surface Pro, a personal phone) give them text and a fleet tracker.

The weekend after the interview, my co-founder Usman and I built the first prototype at NexHacks. The Monday after, I sent the first email to Myron Hughes, Assistant Director at Greene County EMS.

Customer discovery: three days in Greene County

Feb 20-22, 2026 · 4 crew interviews · 1 full shift in MED6

Kitty-Jo drove us up to Greeneville. Over three days we ran four interviews and spent a shift in the truck. What we got:

Hardware reality

What's actually in the truck:

The map tab on tablet hardware that costs more than my rent is a fleet tracker. No navigation, no routing. Every EMT I rode with used Google Maps on a personal phone.

The wedge: landing zones

Greene County has zero FAA-registered helipads. Every helicopter call is an ad-hoc landing: a church parking lot, a school field, a Walmart. Crews spend critical minutes finding a flat 100×100 ft area within a survivable radius of the patient.

Myron flagged this. Kitty-Jo flagged this. Dispatch flagged it. Independently. That was the wedge. Not navigation. Landing zones.

Product decisions, made in the truck

Almost every pre-ride-along plan got revised.

1. Phone-first, not tablet-first

Original plan: Tauri desktop client on the truck's Surface. Reality: that Surface has 4 GB of RAM, can't run Vectr next to IdnMobile (CAD), and crews already don't trust the map tab. Pivot: native iOS with CarPlay support as the primary surface; the Tauri client became a fallback for dispatch desks and outer-zone trucks with phone dead spots.

2. Offline-first

Rural East Tennessee has cell dead zones, specifically the area near the NC line. Hospitals are an hour away. If the network is a hard dependency, the product breaks on the calls that matter most. So: local-first cache, syncing when online, with map tiles + crew notes + LZ candidates packaged per county.

3. LZ as the painkiller

Navigation is a vitamin: Google Maps + CAD already solve "where is it" at 80%. Landing zones are a real, unsolved, expensive problem. We built lzDetection.ts + LZFinderPanel.svelte, fed it FAA NASR data, USGS 3DEP elevation, and Overture density scoring, then trained a DINOv2-Small (22M parameter) satellite scorer on DGX Spark, precomputing 21 candidate landing zones for a county with zero FAA-registered helipads.

4. On-device AI for the Scribe

Patient data is PHI. PHI in the cloud is a HIPAA + ESO compliance tax we couldn't afford as a pre-revenue startup. So the iOS AI Scribe runs entirely on-device via Cactus: Parakeet TDT for streaming STT (4.65% WER), LFM2-350M for structured extraction, silero-VAD for speech detection. ~450 MB resident memory total, monitored under iOS memory pressure.

5. CAD address auto-flow as table stakes

Director T.J. said it on the second call: "It has to be seamless." If a crew has to retype the dispatch address into Vectr, the product is dead on arrival. IdnMobile has no public API, so we built a clipboard-monitor workaround for the Surface and a separate iOS strategy. This, not LZ scoring or the scribe, was the actual gating constraint on the pilot.

Design choices

• Dark theme everywhere. Crews work nights. White UI in a dark cab destroys night vision.
• 24-hour time. Military convention; what every EMS clock and run sheet already uses.
• Map as the primary canvas. Every other surface is overlay. No login screen between dispatch and intel.
• One-handed operation. Buttons sized for gloved hands. Critical actions reachable with a thumb on a phone mounted to the dash.
• "R&D preview" labels on AI scores. LZ quality scores are model output; never presented as ground truth to crews making decisions in seconds.

The pilot path

After the ride-along we drafted a pilot agreement:

• Free 6-month trial on units MED1 and MED6
• Business Associate Agreement for patient data handling
• Cancel anytime with 30 days notice
• Crews retain decision authority (Vectr is decision support, not dispatch)

Sent to Myron and his Director T.J. on March 6, 2026. Went to county legal the next week. Myron presented our one-pager at a regional directors meeting on March 12; a few directors scanned the QR code.

Then it stalled. County legal didn't come back. The directors didn't follow up. By late March I was checking in weekly with no concrete movement.

Why I closed it

I graduated in May 2026. The project had been drifting since March, and the math wasn't there:

• ~7 weeks of silence from Greene County after legal stall
• My co-founder Usman wanted to dissolve the entity
• I'd moved to Virginia and started a different primary project
• The competitor space (RapidSOS, FirstNet, Prepared, Carbyne) is well-capitalized and full-time; the only honest path forward part-time was cold-emailing rural EMS departments without conviction

I sent Myron a wind-down email on May 16, 2026. He'd been one of the most generous customers a first-time founder could ask for; the relationship deserved a clean close, not a slow ghost.

The legal entity (Vectr Systems Inc., Delaware C-corp) stays alive through Greene County's June budget meeting, then dissolves.

What I'd do differently

1. Move on dispatch integration first. Every other product decision was downstream of "can we get the address automatically?" We treated it as a future problem. It was the first problem.
2. Run the entity decision in parallel with the pilot, not after. A 50/50 split with a co-founder who doesn't share your conviction is a slow leak.
3. Don't treat the field partner as a "case study." Treat them as the only thing that matters. Show up more than three days. Eat the dinner. Call between visits.
4. Sell the painkiller, not the vision. The whitepaper pitches Scene Intel + AI Scribe as a system. Greene County wanted the LZ tool that works on the helicopter calls. Should have led with one feature, not five.
5. Closure is a feature. Closing well let me sleep, kept the relationships intact, and made the work portable into this case study. Not every project ships. Every project should end on purpose.