Physician-founder and CEO of DNAi Systems. I build Asha, a retrieval-grounded fiduciary medical AI, and the platform it runs on.
Asha is a neurosymbolic medical agent. The language model is the verbalization layer; the intelligence sits around it: KIL retrieval over a curated Qdrant corpus (144M+ vectors, ~1,071 collections, ~314 of them knowledge-source), an Epistemic Arena where candidate claims compete before anything is said, and META_CORRECT, a deterministic post-emission corrector. US Patent 12,555,008 B1, granted 2026-02-17.
Every benchmark number DNAi publishes lives with its pipeline, grader, and artifacts at dnai.systems/benchmarks. Three rows from the current register:
- HealthBench Hard, 1,000 items, official gpt-4.1 grader: Asha 42.0 against 29.4 for its own bare composer (GPT-6 Astra), +12.6 pp paired on the same items (run 2026-09-21).
- MedQA USMLE, 1,273 locked items, single shot: 95.52%, 0 parse failures (2026-05-04).
- psychosis-bench, 96 intervention-eligible turns: 95.8% safe-intervention rate against 30.2% for the same model family (Gemini 2.5) without Asha (2026-05-11).
asha-benchmarks: every public benchmark artifact in one place. The same language model alone and inside Asha, on the same items under the same grader: HealthBench (all 5,000, Hard, Professional, the Vishwanath et al. 500), MedQA, psychosis-bench, enclomiphene-bench and a locked pediatric chart, with per-item data, SHA-256 manifests, pre-registrations, model cards and A2A agent cards. Live register: dnai.systems/benchmarks.
The Citadel monorepo (FastAPI, React + Vite, Qdrant, PostgreSQL, Redis, Docker; about 3,200 tracked files and 3,000 commits on main since December 2025, 97% of them by the two physician co-founders) is private because it runs a live medical product with real users. Walkthrough on request.
Co-founder: Paridhi Anand, MD.
Contact: dsingh@dnai.systems