IntuitAI
Agentic AI venture studio
We build the reliability layer that multi-agent AI is missing — and turn it into companies.
Agentic systems fail for boring reasons. Two agents wait on each other for ever. A loop runs a hundred iterations without making progress. One agent trusts a peer that has been wrong all week, and nobody can say afterwards which one caused the bad answer. Those are coordination problems, not intelligence problems — which makes them tractable, testable, and worth building a company around.
IntuitAI researches those failure modes, publishes reference implementations in the open, and incubates the ones that find real users into focused startups.
- 13open-source repositories
- 9with working code
- Go + Pythonprimary languages
- Apache 2.0licence on all 13
Investment thesis
The industry spent three years making individual models better and has spent the last one discovering that a system of agents is not the sum of its models. The bottleneck has moved: from what a model can do in a single turn, to what a fleet of them does over hours, across tools, under partial failure.
That shift creates an infrastructure gap with the shape investors recognise — the same shape distributed systems had in 2012, before tracing, circuit breaking and service meshes became products. Multi-agent deployments today have no equivalent of a deadlock detector, no calculus for how much one agent should trust another, no attribution trail from a wrong answer back to the agent that caused it, and no cache that understands that yesterday's answer is stale because the underlying document changed.
We are building those primitives one at a time, in public, as working software rather than position papers. Each is a plausible company on its own; together they are a coherent stack, and a studio holding several of them is better positioned than one betting on any single primitive being the winner.
How we incubate
The studio runs one pipeline, and a project only advances when it clears the gate in front of it. Nothing graduates on enthusiasm.
-
Characterise the failure
Start from a defect people actually hit, reproduced under test, not from a market map. Tangle exists because agent graphs deadlock; MultiTrust exists because a scalar reliability score cannot express "I have not seen enough of this agent to judge it."
-
Build a reference implementation in the open
Apache 2.0, CI, and coverage from the first commit. Publishing the reference implementation is how the studio gets its evidence: adoption of something free is a cheaper and more honest demand signal than a survey.
-
Test the demand, not the demo
Integrations with the frameworks teams already run — LangGraph, OpenTelemetry, MCP, the OpenAI API surface — so that trying a project costs an afternoon rather than a migration. Usage, issues and integration requests are the gate.
-
Spin out under stealth
A project that clears the gate becomes a separate venture with its own team and its own commercial surface, while the reference implementation stays open. Ventures currently in build are unannounced.
How we work
The studio's advantage is not speed of shipping; it is that its claims survive contact with a reviewer. Four habits do most of that work, and they are visible in the repositories rather than asserted here.
- Determinism is a test, not an aspiration. llmsim gates every one of its four parallelism tiers on same-seed determinism: the same master seed must produce identical results on any backend, any worker count, any Python build. A parallel simulator that cannot promise that is not usable as evidence for anything.
- The maths comes before the API. MultiTrust models inter-agent trust as a Subjective Logic opinion — belief, disbelief and uncertainty as three separate quantities — because a single 0-to-1 score silently conflates "this agent is unreliable" with "I have no evidence either way." The formalism is the differentiator; the surface area stays small.
- Failures are asserted loudly. Tangle ships with a 98% test-coverage badge and detectors for wait-for cycles, resource cycles and livelocks, because a monitor that misses the condition it exists to catch is worse than no monitor.
- We publish the limitations with the result. Our reference implementation of the 2025 Duan et al. shortest-path algorithm says in its own README that it does not reproduce the paper's O(m log2/3 n) bound, and lists the three substitutions that cost it. Work that overstates itself is worth less than work that doesn't, and it is easier to fund the second kind.
Incubator projects
All thirteen are public and Apache 2.0. Maturity is labelled, and the labels are meant to be checkable: Shipping means working code under CI, Active means a working core under development, and Design stage means a repository holding a problem statement and not yet a system. Nine of the thirteen have working code; four do not, and say so.
Agent reliability and trust
-
Tangle Shipping
Deadlock and livelock detection for multi-agent workflows. Watches agent interactions in real time, distinguishes an agent that is stuck from one looping without progress, and fires configurable resolution actions. Runs embedded as a Python library or as a standalone FastAPI sidecar, with LangGraph decorators and language-agnostic OpenTelemetry ingest.
-
MultiTrust Shipping
A trust framework SDK for multi-agent systems, grounded in Subjective Logic. Trust between agents is an opinion — belief, disbelief, uncertainty — rather than a score, so a system can act on the difference between distrust and ignorance. Zero hard runtime dependencies; LangGraph, OpenAI, Anthropic, MCP, Redis and OpenTelemetry live behind extras.
-
Oculon Design stage
The thread: observability and attribution for agent fleets — tracing a wrong or expensive outcome back to the specific agent, tool call and context that produced it. Intended as the complement to Tangle, which catches a workflow that has stopped rather than one that finished badly. The repository currently holds the problem statement, not a system.
-
Cogmux Design stage
The thread: a cognitive multiplexer routing work across slow, fast and reactive agents, on the premise that most agent latency and cost is deliberative reasoning aimed at problems a reflex would have solved. A premise we intend to test, not a result we have.
LLM infrastructure
-
Reverb Shipping
A two-tier semantic response cache with knowledge-aware invalidation. Tier one is an exact SHA-256 match; tier two is embedding cosine similarity. Its distinguishing feature is that it tracks which source documents fed each cached answer and invalidates on change-data-capture from those sources, so the cache goes stale when the knowledge does. Ships as a Go library, an HTTP and gRPC service, and a drop-in reverse proxy for any OpenAI-compatible upstream, with opt-in PII redaction before hashing.
-
Random Number MCP Server Active
An MCP server that seeds random-number generation from national weather data. Small by design, and the studio's working reference for shipping, containerising and testing a Model Context Protocol server end to end.
Simulation and workflow research
-
llmsim Active
A parallel discrete-event simulation library for Python 3.14+. Keeps the generator-as-process model that made SimPy usable and rebuilds the rest for free-threaded Python: parallel Monte Carlo replications, conservative single-run PDES with model-provided lookahead, in-run compute offload, and a master-seed tree that makes every result reproducible across backends. The studio uses it to model agent fleets before building them.
-
Agentic Workflow Design stage
Working notes on agent architecture — where a tool-calling agent genuinely beats a chatbot, and what state management, validation and error handling a reliable one actually needs. The written form of the studio's house position on agent design; there is nothing here to install.
Applied systems and engineering foundations
-
RAGsearch Active
A private retrieval-augmented search assistant: PDFs chunked, embedded and indexed into Qdrant, queried semantically, answered by a local model through Ollama. Runs entirely on your own machine — no document leaves it — which is the deployment shape the studio wanted to understand before proposing it to anyone.
-
QuranLLM Design stage
Planned: LLM-powered semantic search over the Quran. Chosen as a deliberate stress test rather than a product — a domain where a confident wrong answer is unacceptable is the right place to work on retrieval grounding and refusal behaviour.
-
resilience4py Shipping
A Python port of resilience4j: circuit breaker, rate limiter, retry and bulkhead as composable decorators, async-native and modular. The fault tolerance patterns that made microservices survivable, in the language most agent code is written in.
-
claude-tools Active
The studio's internal coding-agent skills, published: automated code review, design review, security scanning and documentation generation. How a small team runs several projects at this depth.
-
BMSSP shortest paths Active
A reference implementation of the recursive partitioning structure from Duan et al. (2025), the first deterministic algorithm to break the sorting barrier for directed real-weighted shortest paths. Written for understanding and as a correctness oracle; its README states plainly which substitutions keep it from reaching the paper's asymptotic bound.
Ventures in stealth
Projects that clear the demand gate are spun out as separate ventures. Those currently in build are unannounced — no names, no numbers, and nothing on this page should be read as a claim about their traction. The reference implementations they grew out of stay open source and stay listed above.
Founders, operators and investors who want detail under NDA should get in touch.
Research notes
We publish what we learn, including the parts that did not work.
- ML-powered text recovery from a legacy ANSI-encoded PDF — recovering Bengali text from a memoir typeset in the proprietary SutonnyMJ ANSI font, by treating glyph recognition as an image classification problem with Tesseract 5.x's LSTM OCR, at an estimated 98–99% character accuracy — estimated from sampled pages, not measured against a held-out set, and the note says so. The point of the write-up is the decision, not the tool: reverse-engineering the ANSI-to-Unicode mapping was the obvious route and was intractable, and sidestepping it produced the clean UTF-8 text now published in the Internet Archive edition.
Work with us
Investors. We are happy to walk through the pipeline, the gate criteria and what is behind the stealth line. Email nobel@outlook.com.
Founders and operators. If you are working on agent reliability, inter-agent trust, or inference efficiency and want a studio rather than a term sheet, the same address reaches us.
Engineers. Every project above takes issues and pull requests. Start at github.com/nobelk, where the repositories live, or at the IntuitAI organization.
IntuitAI is led by Dr. Nobel Khandaker, whose doctoral research was in multi-agent systems.
Source code · Machine learning icons created by Becris — Flaticon