We build foundational AI technology — open standards for evidence integrity, grammar-first tokenizers for complex languages, absence-detection architectures for safety-critical systems, and voice identity verification infrastructure. Published on arXiv. Built from first principles.
AI generates convincing fake photos in seconds. EIF is an open standard that evaluates whether a photograph constitutes reliable evidence for a specific claim — not just "is this real?" but "is this trustworthy?"
C2PA tracks provenance — who created it, what device, what edits. Deepfake detectors classify real or fake. Neither evaluates whether a photo is reliable evidence for a specific claim.
EIF analyses the photograph itself: compression artifacts, noise patterns, lighting physics, and statistical signatures that separate real sensors from neural networks.
A standard for truth cannot itself be opaque. EIF is open.
Patent pending — EIF multi-dimensional metric framework for domain-specific evidence integrity evaluation.
BPE shatters agglutinative words into meaningless bytes. VerChol's grammar-first approach decomposes words at morpheme boundaries — preserving grammatical meaning for 500M+ speakers globally.
Existing tokenizers — SentencePiece, BPE, WordPiece — were designed for isolating languages like English. They systematically fail on languages where a single word carries clause-level meaning through grammatical suffixing.
VerChol's grammar-first approach achieves a 3.1% fertility improvement over BPE on agglutinative language benchmarks — by decomposing words into grammatically meaningful morphemes instead of statistically frequent byte-pairs. Published on arXiv.
BharatMini — Low-Cost Domain Training
Alongside the tokenizer, we demonstrated narrow-domain model training at ₹2,700 — proving domain-specific AI for manufacturing and robotics doesn't require massive compute.
Every AI safety framework detects what IS present. SenseAi inverts this: it detects what SHOULD be present but ISN'T. A fundamentally different computational problem.
The architecture uses a four-state processing model that classifies signal streams by the absence of expected patterns. In autism: the absence of expected physiological variability predicts crisis. In manufacturing: a missing sensor reading means failure.
Applications
Two API calls. Enroll a voiceprint. Verify the speaker. Three layers of defense — identity matching, deepfake detection, and behavioral liveness. What you build with it is up to you.
Three independent layers
1Identity — ECAPA-TDNN voiceprint matching. Language-agnostic, works across accents. 192-dimensional voice embedding.
2Authenticity — AASIST-based deepfake detection. Catches voice cloning, text-to-speech, replay attacks, and AI-generated audio.
3Liveness — Behavioral analysis. Breathing patterns, micro-hesitations, response latency — natural speech characteristics synthetic speech can't replicate.
Any program, in any language, that can read and write files can now use AI. No SDKs. No API keys. No cloud dependency.
Speech-to-text, audio classification, sound event detection. Write audio data to a file — get intelligence back.
/dev/ai/hearImage classification, object detection, document understanding. Any camera, any image, any format.
/dev/ai/seeSummarization, Q&A, translation, analysis. Local models by default. Cloud when you choose.
/dev/ai/thinkSchool dropout. 20+ years in technology. 5+ years in textiles and manufacturing. Worked across finance, trading, insurance, aviation, and ecommerce domains. Solo technical founder — all architecture designed and built personally. Published researcher on arXiv. 4 patents in preparation. Based in Bengaluru, India.
We build from first principles and publish our research. Open standards, open code, open papers.