We pointed our own AI at our own NSF SBIR.

Sureel Ventures LLC is an AI software company that designs, builds, and operates production AI inside regulated industries. It is a for-profit US small business.

Sureel's core technology is a runtime verification and grounding layer that wraps a large-language-model agent operating the live front line of a regulated business (answering phones, taking intake, answering questions from operational data). It provides two properties current conversational AI cannot guarantee: (1) verifiable policy-boundedness — every candidate utterance is checked against an explicit, machine-checkable policy BEFORE it is spoken, so violations are suppressed pre-utterance rather than detected after harm; (2) verifiable grounding with tamper-evident provenance — every factual answer must be entailed by an authorized data source and carries a cryptographic, hash-chained record of what was said, why it was allowed, and which source row it came from.

Sureel already operates a HIPAA-aligned, voice-based clinic-intake agent running on live patient lines, with grounding, policy-routing, and tamper-evident audit-logging in early production form. This gives the project real deployment data and live regulatory constraints to test against, not a synthetic benchmark.

The architecture is a neuro-symbolic split: the LLM proposes a candidate utterance; a symbolic policy verifier and an entailment-based grounding checker dispose. Today's LLM guardrails are probabilistic and prompt-based — they reduce bad outputs but cannot bound them and have no rigorous notion of 'grounded.' Driving violation and ungrounded-answer rates toward zero in unconstrained dialogue, without over-blocking valid speech or exceeding a conversational latency budget, is an open problem.

Phase I builds the verification layer and measures whether it answers the research question. Objective 1: a machine-checkable declarative policy schema for healthcare-intake dialogue (delivered: schema + reference clinic-intake policy; measured: coverage on a labeled corpus of real intake turns). Objective 2: the runtime pre-utterance policy verifier plus an entailment-based grounding verifier emitting the tamper-evident provenance record (measured: precision and recall on held-out labeled turns). Objective 3: benchmark the wrapped agent against two baselines — an unguarded LLM agent and a prompt-only-guardrail agent — measuring policy-violation rate, hallucination/ungrounded-answer rate, task-completion rate, and added latency and per-conversation cost. Objective 4: mid-conversation routing that classifies in real time whether a request must be handled by BAA-covered infrastructure or escalated to a human, without leaking what the caller disclosed. The go/no-go gate for Phase II is a defensible measured reduction in violation + ungrounded rates over both baselines at acceptable task-completion and latency.

The immediate customer is the small-to-mid healthcare practice — clinics and telehealth providers drowning in phone, intake, and after-hours demand who cannot risk an AI that says the wrong thing under HIPAA. They already pay for answering services and after-hours staff. The opportunity expands horizontally to other regulated small businesses (financial services, legal, government-facing operations) where the verification machinery is identical and only the policy specification changes, and up-market to mid-market and enterprise regulated organizations with larger liability and compliance budgets. Because the Phase-I result is a domain-agnostic verification API, it generalizes into a licensable trust layer other platforms embed.

The broader societal impact is healthcare access: a trustworthy intake agent extends a small practice's reach (after-hours scheduling, fewer missed calls, faster triage routing) without adding staff — disproportionately valuable for under-resourced and rural providers who cannot afford 24/7 front-desk coverage and where a missed call can mean a missed visit.

Doug Waun is Founder & CEO and the Principal Investigator; he has built and runs production AI across multiple regulated businesses and is already at least 50% employed by Sureel, meeting NSF's PI rule. Mike Ion, PhD (Mathematics) is a senior technical advisor (consultant) bringing AI/LLM research depth for the neuro-symbolic verification and entailment-grounding work. The company's edge is the combination NSF prizes: a real in-production commercial wedge grounded in regulatory data, paired with PhD-grade research depth and the discipline to measure honestly, including reporting a negative result if the verifier cannot beat existing guardrails on the latency/quality trade-off.

Phase I PRODUCES the violation-rate, ungrounded-rate, task-completion, and latency numbers — the company does not yet have measured Phase-I benchmark results and does not claim any. The methodology is the deliverable; the metrics are the output of the work, not an assumption.