Venkata Sai Pranav Regulagedda · KIT / Max Planck Institute for Plasma Physics · KIT Gründerschmiede · July 2026
Remember everything, or respect everyone?
Forget
A conference. A handshake. A face you know you've met.
…what was his name?
Your context is not just digital. It's the room you're in, the label you read, the person in front of you. Today's AI is blind to all of it.
Record
7 million AI camera glasses sold in 2025, triple the year before.EssilorLuxottica · CNBC 2026
Humane's always-on AI Pin: shut down; assets sold to HP for $116M.TechCrunch · Feb 2025
EU bans untargeted facial-image scraping & real-time biometric ID.EU AI Act · in force Feb 2025
1 wearer × 1 waking day of 1080p video ≈ 0 GB = other people's faces, voices and screens, retained on a server, searchable forever.
Both sides lose. The dilemma is architectural: assistants that must store context to use it will always end here.
Perceive, don't record.
Trace is a context-first ambient memory. It turns what the camera sees into typed, cited text evidence, on the device and in real time, then destroys the frame.
Raw pixels live for milliseconds. There is nothing to leak, subpoena, or breach. Surveillance is structurally impossible, not policy promised. Native GDPR fit.
②
Never pretends to know.
Every answer cites its evidence, or refuses.
It cannot violate your privacy, and it cannot lie to you. By architecture, not by promise.
It runs.
It cites.
It refuses.
Working end-to-end prototype: live camera perception to cited answers, entirely on local models. No cloud. No API keys. No frames kept.
① It runs
live capture · iPhone
14:31:58 desk · monitor · mug 14:32:04 black keyboard 14:32:09 label text: "DEWALT" 14:32:15 person entered · left
frames kept: 0
Real-time perception overlay; evidence lines appear as they are seen.
② It cites
Q: Where did I leave the drill?
A: On the garage workbench, next to the red toolbox.
Every answer carries its evidence rows and a calibrated badge.
③ It refuses
Q: How many unicorns are on the desk?
no evidence
Measured: 0% confident-wrong answers on a reproducible benchmark battery.
✓ on-device perception (iPhone, real-time)✓ text-evidence memory, encrypted, never deleted✓ local LLM answering, no cloud✓ 240+ automated tests✓ one-command reproducible benchmark
Built solo, on consumer hardware.
From research prototype to Europe's memory layer.
HORIZON 1 · now → 6 mo
Harden
R&DCounting precision: today Trace answers "between 2 and 3" rather than guess. Goal: firm counts without sacrificing 0% confident-wrong · multi-room memory · 10 daily-use testers
MarketLanding page + waitlist · privacy / local-AI community presence · first demo videos
HORIZON 2 · 6 → 12 mo
EXIST phase
R&DWearable form factor: capture moves from iPhone to the glasses-clip charm · closed beta, 100 users
MarketCo-founder onboard · brand + content engine · IP & open-source licensing settled · first design-partner talks (EU device makers)
HORIZON 3 · 12 → 24 mo
Ship
R&DPublic product, multi-device
MarketFirst revenue · seed round raised on beta traction
Budget philosophy: stipend-phase spend concentrates on R&D and marketing. No offices, no overhead theater.
Built by one person. Designed for three.
Venkata Sai Pranav Regulagedda
Founder · KIT · thesis at Max Planck Institute for Plasma Physics
Materials scientist by training, full-stack AI builder by proof: iOS capture app, real-time on-device perception, encrypted memory store, local-LLM answering, and the reproducible honesty benchmark. Every layer, one person.
"I didn't come from computer science. I came from measurement science. That's why Trace refuses to guess."
① iOS capture app
② real-time perception
③ encrypted memory store
④ local-LLM answering
⑤ honesty benchmark
Research collaborator
Latheesh Roy · CSE, remote. Works under Pranav's mentorship on the 0%-hallucination research line: refusal calibration and the benchmark battery.
Deliberately not a co-founder tile. The honest label.
Product / ML Engineer
Co-founder seat: open. Owns perception R&D and the wearable port. Sought through KIT's engineering and Gründerschmiede network.
EXIST eligibility: KIT-affiliated founder · team completion is an explicit Horizon-2 milestone.
Every "personalized" AI answer needs physical context. Only Trace can supply it without surveillance.
TAM · AI smart glasses / wearables≈ €8 to 15B by 2030
SAM · EU/DACH privacy segment≈ €2B · the harder the law, the bigger our edge
SOMDACH early adopters · bottom-up ↓
100k reachable enthusiasts × 15% × €5/mo × 12 = €0.0M ARRwithin 24 mo of launch
Stage 2 · the platform market: the context layer for all AI
Siri, Gemini, ChatGPT: every assistant is racing toward personalized answers, and personalization needs your physical life. None can capture it without becoming a surveillance product.
Trace as licensed context layer: >2B phones already run local AI. Even €1/user/mo across a fraction of the billions of assistant endpoints is a multi-€B annual platform market. Wearables are the wedge.
Why now · three converging curves
1 · On-device AI crossed the threshold. >2B phones now run local AI; sub-20ms on-device vision. Trace was impossible 24 months ago.
2 · AI wearables ship in volume. 7M glasses sold in 2025, while the pure recorders (Humane, Limitless) died at the privacy wall.
3 · EU AI Act is live. Biometric and facial-scraping bans since Feb 2025. "Structurally non-retaining" becomes procurement criteria.
TAM: Grand View Research ($8.26B smart glasses by 2030) · IDC (installed base >80M units). SAM/SOM are stated bottom-up assumptions. Full assumption table in appendix; every number carries a source or a stated assumption.
Everyone else must choose: see your world, or respect it.
Physical-world context →
← Digital-only
Zero retention ↑
↓ Raw data retained
Meta Ray-Ban / AI glasses
Rewind · Limitless
Apple Intelligence · Gemini
Notion / PKM tools
TRACE
Physical context has always required retention. That's why this quadrant is empty. The two purpose-built recorders of this generation, Humane and Limitless, are already dead or absorbed. Trace's architecture is the first way in.
Why won't Apple or Google just build this?
Their business models monetize retained data and engagement; their privacy stack deliberately stops at your digital life. A zero-retention physical layer contradicts their data economics. If the giants want this capability, they license it or acquire it. Both outcomes require Trace to exist first.
The device wins adoption.The layer earns the money.
Stream 1 · Consumer · adoption engine
Trace Membership · priced to spread
One low monthly price (target €4 to 6/mo), charm included. Clips to your existing glasses; seasonal editions keep it a fashion object, not a gadget.
BOM ~€15 to 25 at volume (display-less: Ray-Ban's costliest part, >50% of its BOM, removed) → membership covers hardware and logistics → consumer margin ≈ 0, by design
marginal compute cost ≈ €0. The AI runs on the customer's phone. We can afford adoption pricing forever; cloud-AI competitors structurally cannot.
Stream 2 · B2B licensing · profit engine
The context layer, licensed per device
To eyewear brands, wearable makers, and assistant platforms that need physical context without surveillance liability. Software licensing at ~100% gross margin.
Every membership user is proof, benchmark data, and negotiating leverage for this stream.
Arm (~$1.8B/yr) doesn't sell chips; Dolby (~$1.2B/yr) doesn't sell speakers. Each licenses a layer into devices it doesn't build. Trace doesn't need to profit on charms; it needs to become the layer every device licenses.
Price points are validation targets for the EXIST phase, not commitments. Consumer pricing research is an explicit Horizon-1 milestone.
Four asks. Each one unblocks a milestone.
1 · EXIST-Gründerstipendium application prep
funds Horizons 1 and 2: counting-precision R&D, the wearable charm port, the first marketing engine