Baljit Singh Brain-inspired AI, memory-first systems, and hardware-aware architecture

Personal site of Baljit Singh • Founder, Mankash AI Labs • Available for selective consulting

Building connected intelligence with memory, structure, correction, and hardware-aware architecture.

I work at the intersection of brain-inspired AI, neuro-symbolic systems, streaming language understanding, memory-centric architectures, LLMs, multi-agent systems, and chip design thinking.

I help serious AI teams design systems that move beyond demos into reliable, inspectable, and differentiated production architecture. My north star: intelligence is not only prediction. It is connectedness, correction, durable memory, and architecture that can live in the real world.

Explore consulting See what I have built Read the blog

Selective consulting availability: 40 to 80 hours per month • Remote with US and global teams • Architecture reviews, technical sprints, and advisory retainers

Explore consulting →
Best fit Startups, CTOs, advanced R&D teams, and enterprise AI groups building differentiated systems.
Core lens Memory, symbolic structure, correction loops, and hardware-aware system design.
Consulting formats Architecture diagnostics, focused technical sprints, and fractional AI architect retainers.
Operating range LLMs, RAG, agents, neuro-symbolic pipelines, speech, healthcare AI, and AI hardware tradeoffs.

Perspective

A modern AI thesis with a hardware conscience.

I am interested in AI systems where architecture matters as much as optimization. That means topology, neurmorphology, symbolic structure, memory, and correction loops are part of the design from the start rather than retrofits.

Structure before scale

Topology and representation design are not containers for intelligence. They are part of its source.

Memory as architecture

Long-term, editable, inspectable memory should shape system behavior instead of sitting outside it. This is also where I help teams rethink RAG, long-term memory, and context systems that must survive real workflows.

Corrective intelligence

The systems I care about should revise, reconcile, and repair their internal state as new evidence arrives. For production AI, correction is the difference between one-shot answers and systems that can be reviewed, repaired, and improved.

Compute substrate matters

Latency, memory movement, precision, noise, and analog behavior all shape what kind of AI is feasible.

Systems

What I have built, and what those systems taught me.

The projects below are best understood as architectural experiments in language, structure, perception, memory, and explainability.

Neuro-symbolic architecture

GIPCA and BISLU

I worked on architectures that merge statistical AI and symbolic AI by forcing meaningful, inspectable boundaries between stages. This supports explainability and stronger learning behavior in settings where black-box models are not enough.

This direction is embodied in GIPCA (General Intelligence Predictive and Corrective Architecture) and BISLU (Brain Inspired Spoken Language Understanding).

Universal NLU

Thought-level outputs, not only intent labels

Instead of compressing meaning into small intent sets, I worked toward human thought representations with variable resolution depending on domain confidence.

Streaming speech

Meaning accumulation and Centom segmentation

Real speech is fragmented and messy. My work on streaming NLU and Centom focused on segmenting spoken streams using atomic connected entities while preserving local meaning.

Thought representation

ETML and symbolic thought clouds

I built toward graphical thought representations for conversation and context, plus ETML as a textual form engineers can inspect, debug, and evolve.

Speech + perception

Medical-ready STT and visual de-referencing

I worked on lightweight speech-to-text for noisy medical settings and on visual context enrichment so gestures can contribute meaning to language understanding.

Brain-inspired simulation

Spiking neural systems and neocortex simulation

I have studied spiking neural networks and simulated behavior from real neuron morphology to understand how perceptual features correlate across connected structures.

Proof

Patents, hardware depth, and a broad systems range.

My differentiator is not just one model or one product category. It is the ability to connect language, AI architecture, memory, hardware, and execution into one practical view.

Selected patents
  • Automated system for digitization and analysis of handwritten medical records
  • Brain Inspired Spoken Language Understanding System (BISLU)
  • GIPCA based system for inferring phonetic based words
  • GIPCA based system to convert thought representation into coherent stories
  • GIPCA based system for inferring human thought representations
  • Intelligent footfall analysis in hospitals
  • Revenue leakage detection in hospitals
  • Inferencing adverse health conditions of a patient
Artificial intelligence illustration Hardware and chip design

I have worked across hardware development and chip design alongside software and AI systems. That perspective keeps questions of memory bandwidth, precision, power, noise tolerance, and analog directions in the room from day one.

Deep Learning Hardware Development Chip Design Neuro-Symbolic AI LLMs Multi-Agent Systems Spiking Neural Networks Brain Topology Analog AI

Next

What I want to build next.

These are collaboration-ready directions where I believe a strong architecture point of view can create disproportionate results.

Symbolic RAG with thought representations

Retrieval grounded in thought structures rather than only text chunks, with cleaner provenance and better debuggability.

Memory-native LLM systems

Persistent, editable memory that changes behavior over time and supports correction, provenance, and long-lived context.

Analog and corrective AI architectures

Noise-tolerant compute and hardware-software co-design where approximation and correction are normal parts of robust intelligence.

Consulting

Architecture help for teams building beyond the demo stage.

Many AI products look impressive in a controlled demo but struggle when memory, reliability, provenance, latency, cost, and real user workflows enter the system. I help teams reason through those architecture gaps and design systems that are easier to debug, improve, and trust.

Diagnostic

AI Architecture Diagnostic

A focused review of your RAG, agent, memory, evaluation, cost, and reliability architecture, with a practical roadmap before you scale, rebuild, or raise.

Sprint

Memory-Native LLM and Symbolic RAG Sprint

A design sprint for systems that need structured memory, provenance, thought-level representations, correction loops, and better retrieval behavior.

Retainer

Fractional AI Architect

Ongoing 40 to 80 hour monthly advisory for founders and CTOs who need senior technical judgment without a full-time executive hire.

See consulting options Compare engagement formats

Collaborate

If you are building differentiated AI, I would like to compare notes.

The strongest conversations begin with a real architecture constraint. Send what you are building, where the current system is failing or uncertain, and what success needs to look like.

  • What are you building?
  • What is the hard architecture question?
  • What constraints matter most: accuracy, latency, cost, memory, auditability, privacy, or scale?
  • Are you looking for a diagnostic, sprint, or ongoing advisory support?