KBKim Boher

Ventures & projects · technical case files

Built before. Shipped before.

Production ventures across forensic AI, computer vision, agentic systems, blockchain evidence, voice AI, TinyML, and full-stack products. The descriptions below preserve the original project record; implementation evidence and demos open inside each case file.

ScopeFounder · CTO · BuilderEvidenceLive products · demos · field records
01

Founder & CEO — forensic fraud detection · U.S. patent pending

DeFake — Forensic AI Fraud Detection

A patent-pending six-layer forensic waterfall checks media integrity, origin and provenance, tampering, metadata, hardware fingerprints, watermarks, and neural artifacts across video, images, and voice. It reaches up to 98.9% detection accuracy across 11 AI video generators, then produces the evidence each workflow needs: forensic reports, court evidence, financial-damage evaluations, takedown notices, and verdicts.

Detects AI-generated, tampered, manipulated, and staged content across video, images, and voice — then produces the evidence and next action.

  • Forensic AI across video, images & voice
  • 98.9% across 11 AI generators
  • U.S. patent pending
  • Live product
Visit DeFake ↗Read the benchmark whitepaper ↗
Evidence, implementation & demos
Suspect media inDeFake analysisProof, delivered
1Fraudulent insurance claim2Stolen identity video3Fake news footage4Deepfaked court evidence5Synthetic KYC identity
DEFAKE
Media integrityOrigin & provenanceTamper analysis
1Forensic report2Court evidence3Financial damage evaluation4Takedown notice5Verdict
Suspect media in · forensic proof outLIVE on defakes.com
  • Forensic detection across video, images, and voice — including AI-generated, tampered, manipulated, and staged content
  • A six-layer waterfall combines C2PA provenance, chain of custody, EXIF and hardware fingerprints, watermarks, neural-artifact analysis, and traces of the models or programs used
  • Actionable outputs for each vertical: forensic reports, takedown notices, insurance evidence packages, and expert-review workflows
  • Patent-pending multi-layer waterfall architecture with blockchain-anchored evidence preservation
02

Founder — patent-pending human-in-the-loop data system

DeFake Game — Gamifying Forensic Science

A gamified science and data-acquisition system: users analyze uploaded footage or social links, publish checks to a shared feed, vote AI or real, explain their reasoning, and earn credits for correct answers. Behind the interface, those interactions produce structured labels and heatmaps for model evaluation and training — part of DeFake’s patent-pending technology.

Upload footage or paste a social-media link, test whether it is AI-generated, publish the result, and join the community vote.

  • Human-in-the-loop labeling
  • Heatmaps
  • Community consensus
  • Patent-pending technology
Play the DeFake Game ↗
Evidence, implementation & demos
DEFAKE / COMMUNITY LABCredits · 240
SOCIAL LINK INGESTEDAnalyzing footage…
YOUR VERDICT
REALAI

Temporal artifacts around face and hands

+20 credits · correct answer
01 · Upload or paste a link02 · Vote AI or real03 · Explain the signal04 · Labels + heatmap created
  • Upload footage or paste a link from any social platform to test whether the content is AI-generated
  • Publish analyzed media into a shared feed where other users vote AI or real and explain the signals behind their decision
  • Community interactions produce structured data labels and visual heatmaps in the background
  • Correct answers earn credits that users can apply toward DeFake services
  • Gamifies scientific participation while building human-in-the-loop evidence for model evaluation and training
03

DeFake vertical — damage disputes settled on evidence

Claim Resolution for Car Rental

A mobile-capture and forensic-comparison pipeline guides pickup and return scans, verifies full vehicle coverage, checks every clip for integrity, detects before/after damage, and assembles a claim-ready evidence report.

Scan the car with a mobile phone — damage disputes settle on evidence, not arguments.

  • Built on the DeFake engine
  • 3D capture + forensics
  • Live product
Open live product ↗
Evidence, implementation & demos
✓ Front✓ Rear✓ Left side✓ Right side
Renter
Rental service
Start of the tripEnd of the trip
Full-car coverage✓Identity continuity✓Media integrity✓Before / after comparison✓Staged-footage check✓
Damage identified · left front bumper
Verdict — new damage confirmed · claim supported
$500
Renter · pickup

At pickup, the renter scans the car — every side.

  • Guided vehicle scans at pickup and return, with full-coverage verification
  • Before/after damage comparison plus media integrity checks on every clip
  • Automatic claim-ready report — pickup-to-payout evidence for rental platforms
04

Founder — computer vision & 3D reconstruction

AR Digital Twin

A complete computer-vision pipeline starts with dataset creation and ordinary mobile footage, trains neural networks for NeRF reconstruction, extracts 3D objects and scenes, and serves them in production AR and VR experiences — roughly 30 minutes of compute instead of weeks of manual modeling.

Plain phone video into a production-ready 3D model in 30 minutes of compute.

  • NVIDIA Inception
  • Microsoft for Startups
  • AWS for Startups
  • NeRF · PyTorch · CUDA
Visit AR Digital Twin ↗
Open full neural-network training pipeline & demos
  • Replaces $60–$50,000 and up to two weeks of manual 3D modeling per object
  • Dataset creation, guided mobile capture, neural-network training, NeRF reconstruction, 3D object and scene extraction, and production XR serving
  • The same technology a San Francisco startup later raised $25M to commercialize
  • Deployed with real clients: custom-furniture buyers preview pieces in AR at home before purchase
05

Hardware to model to deployment — solo (HarvardX)

TinyML on Microcontrollers

A complete embedded-ML pipeline built for 256 KB of RAM: dataset creation, model training and evaluation, distillation and INT8 quantization, followed by TensorFlow Lite Micro deployment in C++ and live inference on the microcontroller — no cloud required.

Full-stack ML in 256 KB of RAM — hardware to model to deployment, solo.

  • TensorFlow Lite Micro
  • INT8 quantization
  • Arduino Nano 33 BLE
  • C++ · Mbed OS
Watch training & quantization ↗TinyML at HarvardX ↗
Evidence, implementation & demos
Kim Boher holding an assembled TinyML rig with Arduino Nano 33 BLE Sense
  • Assemble the hardware, create the dataset, train and evaluate the model, then distill and quantize it for the target device
  • Ship it running on the microcontroller itself — no cloud
  • Person detection, gesture recognition, keyword spotting, visual wake words: all live on-device
06

CTO — Telegram mini-app · attested answers · smart-contract payments

Verity — Blockchain Market Research

An end-to-end production blockchain system: a Telegram mini-app collects research answers, a NEAR Shade AI attester evaluates whether quality requirements are met, its verified attestation activates a Rust smart contract, and payment moves automatically into the respondent’s smart wallet.

AI-powered market research for early-stage founders — real users, paid per quality answer.

  • Telegram mini app
  • NEAR Shade agent
  • Rust smart contracts
  • Demo Day
Open live demo ↗
Evidence, implementation & demos
  • Founder describes the product; AI generates a tailored questionnaire, shared by link
  • A NEAR Shade attester agent scores every answer — coherence, relevance, uniqueness, thoughtfulness
  • A verified quality attestation automatically activates the NEAR smart contract and transfers payment between smart wallets
  • Shipped as a Telegram mini app: native auth, smart-wallet integration
  • Stack: React 18 + TypeScript, FastAPI, Supabase (RLS + realtime), Rust smart contracts with staking-based anti-farming
07

Realtime voice pipeline · fine-tuned self-hosted open-source LLM

Voice-to-Voice AI Assistant

Built an end-to-end voice-to-voice sales agent before realtime voice APIs existed: speech-to-text and text-to-speech models were orchestrated around a fine-tuned, self-hosted open-source LLM, with instruction handling, conversational memory, RAG retrieval, streaming, buffering, latency control, session state, and speech-interruption handling.

A virtual voice-to-voice sales rep, hand-assembled before realtime voice APIs existed.

  • Fine-tuned self-hosted LLM
  • Conversational memory + RAG
  • STT + TTS orchestration
  • Streaming + interruption handling
Read the build story ↗
Evidence, implementation & demos
  • Fine-tuned and self-hosted an open-source LLM to follow instructions and drive the reasoning pipeline
  • Built conversational memory and connected RAG retrieval to the sales knowledge base
  • Orchestrated speech-to-text, LLM reasoning, retrieval, and text-to-speech as separate models and services
  • Implemented streaming, buffering, latency control, voice activity, interruption handling, and realtime session state
  • Published the thesis that voice AI would break out in 2025 — before it was consensus
08

Founder build at Maverick Media — campaign intelligence

Analytical Agent for Customer Acquisition

A customer-acquisition intelligence system combines campaign performance, website behavior, cross-channel user journeys, chatbot leads, and phone calls. It measures both quantity and lead quality, transcribes conversations, extracts sentiment, intent, questions, and objections, and feeds those signals back into campaigns to acquire customers more cheaply and increase sales.

A customer-acquisition management system that combines quantitative and qualitative data to reduce acquisition cost and increase sales.

  • Call transcription + sentiment
  • Source attribution
  • Campaign optimization loop
Evidence, implementation & demos
  • Call outcomes — received / missed / failed / busy — with busiest hours and daily distribution
  • Source attribution by social platform and search keyword
  • Website behavior and cross-channel customer journeys connected to chatbot, form, and call outcomes
  • Every call transcribed and analyzed: sentiment, expressed emotions, key questions, key concerns
  • Every metric feeds back into optimizing the client’s next campaign — built before AI agents became a category
09

CTO at Kora Media — 80-store e-commerce portal, team of 6 engineers

E-commerce Product Content Generator

A schema-free Chrome extension understands product-page content without external scraping, analyzes the product image with an AI agent, and generates SEO titles, descriptions, keywords, metadata, and upload-ready JSON — without database integration or image storage.

Builds complete SEO product content with an AI agent — from a product page and image to titles, descriptions, keywords, metadata, and upload-ready JSON.

  • Google AI Developer Competition
  • Chrome extension
  • Schema-free
  • Genkit + Node.js
Official entry on ai.google.dev ↗
Evidence, implementation & demos
  • AI analyzes both the product page and image, then writes SEO-ready titles, descriptions, keywords, and metadata
  • Schema-free by design: real stores don’t share a product schema, so it works on any website — no database integration, no image storage
  • Technical implementation: a Chrome extension captures the product image in-page → Google’s Genkit AI framework + multimodal model API through a thin Node.js relay → structured listing fields assembled into upload-ready JSON on the client
  • Built before AI agents and in-browser AI existed
10

Built at the NEAR Foundation’s Oz City incubator

Cryptographic Chain of Custody

An online-camera evidence system verifies footage at capture, creates its cryptographic record, and anchors that record on NEAR — preserving provenance and making later tampering independently detectable.

Documents online-camera footage on-chain and creates cryptographic proof of evidence that can be independently verified later.

  • Blockchain evidence anchoring
  • On-device capture
  • Tamper-proof records
Evidence, implementation & demos
Kim Boher presenting at the NEAR Foundation Oz City incubator
11

Founder — LLM legal agents

Cognitive Intelligence Labs

A multi-agent legal workflow combining document ingestion, evidence analysis, live knowledge bases, legal research, case-building, and client intake around a legal-trained LLM.

An integrated legal suite powered by a legal-trained LLM and specialized agents.

  • Ministry of Innovation Accelerator
  • LLM agent architecture
  • Demo Day
Evidence, implementation & demos
Kim Boher presenting Cognitive Intelligence Labs at the MAOF accelerator demo day
  • Automated document handling and evidence analysis with real-time knowledge bases
  • Legal research, case building, and a client-side intake chatbot
  • Presented at the MAOF accelerator Demo Day
12

Founder, CEO & CTO — software & digital-marketing agency

Maverick Media

Services included fractional CTO leadership, software development, deep-tech and AI development, analytical infrastructure, SEO, marketing automation, and management of Google PPC and social-media marketing campaigns.

Founded a technology agency focused on building companies and their operating infrastructure from the ground up.

  • Fractional CTO
  • Software + AI development
  • Analytics infrastructure
  • Growth operations
Evidence, implementation & demos
  • Built companies’ technical foundations: product architecture, full-stack software, chatbots, mobile apps, and analytical infrastructure
  • Provided fractional CTO, deep-tech, and AI development services
  • Delivered SEO, marketing automation, Google PPC, social-media campaigns, conversion tracking, and sales handoff workflows