DevHeal is a product company first, but several of the products above exist because of research questions we kept running into. This page lists those research areas honestly — as questions being investigated, not as shipped results.
Problem: Transformer inference cost scales with context length. Hypothesis: hybrid state-space designs can hold long-range context more cheaply for specific workloads. This is the research line behind the NOVA architecture — see Patents & IP.
Problem: pure state-space models trade off some in-context reasoning ability for efficiency. Method: combine state-space layers with knowledge-graph augmentation and template-augmented generation rather than relying on attention alone.
Problem: running useful models on CPU-only hardware without a GPU. Status: NC AI already ships local CPU inference in production — ongoing work is on pushing model size down further without losing usefulness. See Benchmarks.
Problem: agent execution, memory and tool-calling are reimplemented per product today (HiveANT, SwarmOps). Method: extract the common runtime into AgentOS — see Technology.
Problem: general-purpose languages weren't designed with agents and AI invocation as primitives. Result: NC, shipping today, with a dedicated AI bytecode opcode and first-class agent/@tool constructs — documented in our patent filings (see IP page).
Problem: most agent testing is a handful of manual prompts. Hypothesis: scenario-based simulation (tool failures, policy violations, adversarial inputs) catches failure modes unit tests can't. See DevHeal Simulator on the Platform page.
Problem: accuracy benchmarks don't predict production reliability. Method: multi-dimensional scoring across task success, policy compliance, cost and recovery behavior.
Problem: applying SRE discipline to systems whose behavior is partly probabilistic. Status: HiveANT's ACO/ABC swarm coordination and SwarmOps' reinforcement-learned incident response are our current answers, running today.
Problem: preserving business correctness when an agentic workflow fails mid-transaction. Method: checkpoint, state reconciliation, and compensation, generalized from patterns already used inside HiveANT/SwarmOps.
Problem: coordinating many autonomous agents without a central bottleneck. Method: swarm-intelligence coordination (pheromone-graph reinforcement and evaporation) instead of a central scheduler — documented in Patent Family B, see IP page.
Problem: running NC AI on constrained edge/embedded hardware. Status: not yet started — see Edge Runtime on the Technology page.
The research areas above are directions. These are the concrete, named projects behind them today — each with a real result we can point to, not just a hypothesis. Full technical detail, including what each project's own internal review says about its limits, is in the technical white paper.
An orthogonal recurrence core with a gated delta-rule memory bank — O(1) state per step, verified numerically equivalent to a full sequential computation. Trained and evaluated at small scale on CPU; not yet at the parameter scale the design targets.
Read the white paper section →A byte-exact C tokenizer, independently benchmarked against the Unicode UAX #29 grapheme-boundary test suite. Real, working, CPU-benchmarked — by its own documentation, not yet packaged for enterprise use.
Read the white paper section →A continual-learning technique that cuts catastrophic forgetting on a standard benchmark using a few kilobytes of distilled memory instead of storing raw examples. A narrow, specific result — not a claim about general intelligence.
Read the white paper section →Treats an LLM as an untrusted semantic sensor: its output has to pass through typed evidence contracts and a "tool-proof obligation" gate before any tool call is allowed to fire. A working prototype; not yet benchmarked.
Read the white paper section →Earlier-stage work, published for the curious. Real code, no benchmarks (or only self-generated ones) yet — so no performance claims below, just what each one is.
Not every product built on NC is part of the AI Systems Engineering narrative. NeuralEdge is a real, working product — we just don't want it to blur what DevHeal's core platform is for.
A self-contained market analysis and prediction service built with NC — 30+ technical indicators, live market data ingestion, and ensemble ML predictions in a single deployable unit. A proof point that NC can build real domain-specific applications, not a claim about the core engineering platform.