2017
8th grade
The question.
Started asking how a 1.4 kg organ produces memory, curiosity, and self-awareness from nothing but electrical signals. Not the textbook answer. The real mechanism. The thread that ran through every project since.
About Us
We build cortical intelligence.
AI that remembers, sleeps, and knows what it does not know.
Verace AGI is building the architecture that comes after the transformer. A system that remembers in one shot, learns without supervision, consolidates knowledge during offline phases, and monitors its own uncertainty. Our goal is artificial general intelligence grounded in the structure of cognition itself.
The Journey
Nine years from question
to architecture.
2025
AlgoUniversity
Rank 1.
Graphs Camp. Top 50 out of 40,000+ competitors. Graph theory, combinatorics, and the kind of sustained deep work foundational problems demand.
2025
GoQuant, Miami
Millisecond trading.
Back-end C++ for live trading. Real-time market data feeds for Binance and Deribit. First engineering role. First lesson in shipping code where a missed millisecond was money.
2025
Incredible
Production AI.
Shipped an AI platform with 100+ integrations, dual-agent orchestration, real-time voice. Saw every LLM failure mode from the production side. Watched the industry paper over the same holes, month after month.
2025
Amazon
Order Flow Trading.
Wrote and shipped a first-edition book on quantitative finance and order flow. Proof I could take a dense, technical topic and compress it into a coherent framework readers can follow.
Feb 2026
IJTRP
Two papers.
The Artificial Brain: tried to build cognition on top of SLMs. The conclusion was clear. The foundation itself is wrong. StreamRAG v2 shipped the same quarter with sub-millisecond graph latency.
Mar 2026
Founded
Verace AGI.
Left Incredible. Decided no one else was going to build the architecture from scratch. Wrote the first lines of VCI the same week the company was registered. Solo, no fallback.
Now
2026
Training at scale.
Six months in. Cortical architecture built from first principles. Seven pillars integrated into one running system. Seven patents pending. Training at scale as you read this.
2027
Ship
₹50 Cr year.
API launches in H1. First enterprise contracts in Q2. Legal vertical in H2. From a question in 8th grade to revenue in ten years.
The Founder
Krrish Choudhary
Founder & CEO
It started with a question in 8th grade: how does the brain actually work? Not the textbook answer. The real one. How does a 1.4 kg organ produce memory, curiosity, dreams, and self-awareness from nothing but electrical signals? That question never went away.
Working on production AI every day at Incredible, across chat, voice, and agents spanning hundreds of integrations, the cracks became impossible to ignore. Every month brought a new memory hack, a new RAG framework, a new context window extension. All patching the same fundamental problem: transformers don’t remember, don’t learn, and don’t know what they don’t know.
Drafted "The Artificial Brain" in late 2025, published early 2026. An attempt to build a brain-inspired architecture on top of transformers. The result was clear. The transformer structure is wrong. You cannot bolt memory, sleep, and curiosity onto an architecture designed for machine translation in 2017. The foundation itself needs to be replaced.
The realization crystallized. If real intelligence requires cortical columns, local learning, episodic memory, neuromodulation, and consolidation, then someone has to build it from scratch. Not as a research proposal. As working code. Left Incredible in March 2026. Verace AGI was born.
In the months that followed: VCI. A complete cortical column network. One shot episodic memory. Local learning without backprop. Autonomous sleep consolidation. Not a modified transformer. A replacement for it.
Krrish Choudhary
LNMIIT Jaipur · Class of 2027
Track Record
Research, Production Systems & Recognition
Research
The Artificial Brain
The paper that proved transformers are wrong. Attempted a brain-inspired architecture on top of transformers — the result was clear: the foundation itself needs to be replaced. Published IJTRP, Vol. 2, Issue 2, February 2026.
StreamRAG v2
Real-time incremental code graph system with native parser daemons for 8 languages. Sub-0.05ms latency across 20K+ lines. Published IJTRP, February 2026.
Order Flow Trading
Published book on Amazon, First Edition. Quantitative finance and order flow analysis.
Hybrid GraphRAG for Cross-Lingual Legal Citation Retrieval
A multi-signal fusion approach for Swiss legal information systems. Currently in production.
Production AI
Ex Software Developer — Incredible
AI platform integrating 100+ third-party services with custom LLM fine-tuning, dual-agent architecture orchestrating 300+ services, and real-time voice assistant.
Ex Back-End Engineer — GoQuant Technologies, Miami
Low-latency C++ integrations for real-time market data from Binance and Deribit.
The IP Backing
Seven patents pending.
The architectural claims that differentiate VCI from every existing approach are being filed. The moat is at the architecture level, not at the weights level.
Patent 01 / 07
Persistent episodic memory.
One-shot encoding with reliability-scored retrieval. Memory lives in writable structures, not frozen weights.
Filing pending
Patent 02 / 07
Local learning protocols.
Per-column analytical updates. No global gradient graph. Linear cost at scale.
Filing pending
Patent 03 / 07
Offline consolidation.
Autonomous rest-replay-integrate phases. System improves between interactions without new input.
Filing pending
Patent 04 / 07
Uncertainty estimation.
Per-step confidence tracking. The system reports what it does not know and deliberates when unsure.
Filing pending
Patent 05 / 07
Dynamic structural adaptation.
Columns grow, split, merge, and prune at runtime. Architecture reshapes to workload, not fixed at training.
Filing pending
Patent 06 / 07
Protected knowledge retention.
Catastrophic-forgetting prevention. New skills layer on top of existing knowledge, shielded by design.
Filing pending
Patent 07 / 07
Linear associative recall.
Context scales linearly with length via associative memory. No quadratic cost wall, no hard cutoff.
Filing pending
note · each filing corresponds to a mechanism live in VCI today.
Principles
The Verace manifesto.
Four commitments. Every decision at Verace AGI is downstream of these.
Principle 01
Research first. Code second.
Verace AGI
We start from first-principles arguments and published research. Every architectural choice in VCI traces to a paper or a peer-reviewable claim. No design is a hunch. If we cannot argue why it should work before we build it, we do not build it.
Principle 02
Architecture beats scale.
Verace AGI
Scaling transformers will not reach AGI. The limit is not compute. The limit is that the architecture was engineered for machine translation in 2017. More parameters of the same thing do not produce memory, reasoning, or self-awareness. A different substrate does. LeCun agrees. He raised roughly ₹8,600 Cr on the same thesis.
Principle 03
Biology is the blueprint.
Verace AGI
Every mechanism in VCI maps to a real structure of cognition. Remove one and a specific capability disappears. We do not invent mechanisms. We port what is already proven to work in the only system we know that reaches general intelligence.
Principle 04
Honest engineering.
Verace AGI
No faked benchmarks. No demos we cannot reproduce. No model sizes we have not actually trained at. If it is shipped, it is shipped. If it is not, we do not pretend. The moat is the architecture. It speaks for itself.
What We Will Not Do
Our refusals.
Every AI company makes promises. These are the commitments we refuse to make. Our anti-promises are stronger than most companies promises.
Refusal 01 / 05
We will not
Train on user data without consent.
Private data stays private. VCI is architected so every memory has provenance. If a customer asks, we can show what the system learned and from whom.
Refusal 02 / 05
We will not
Ship demos we cannot reproduce.
Every claim on this site is defensible. When we show something working, it works. When we do not, we say so.
Refusal 03 / 05
We will not
Publish benchmarks we cannot back.
No cherry-picked evals. No comparison charts tilted by prompt engineering. When we publish numbers, they come from actual runs, on public settings.
Refusal 04 / 05
We will not
Scale without validating the architecture.
Size follows proof. We are not scaling parameters to mask a broken substrate. We are finishing the training run, then scaling.
Refusal 05 / 05
We will not
Raise without a plan to ship.
Every round has a clear use of funds and a ship date. Pre-seed closes the gap from training complete to API live. That is what the round buys.
A Note From The Founder
No marketing team wrote this page. I did. Every line, every claim, every commitment, every number. I mean all of it.
Nine years ago this was a question. Today it is a company, two papers, seven patents pending, and an architecture training at scale. The question never changed. Everything else around it did.
If any of this resonates, reach out. If it does not, I hope at least you enjoyed the read.
Thanks for making it this far.
Krrish Choudhary
Founder & CEO, Verace AGI