gadi cohen
I build production agent systems in healthcare: the parts that make agents trustworthy.
Staff engineer at Granted Health (founding team) · ex-Assort, Blackbird.AI · Brown CS
Ask my work anything
Answers come from my real work.
Granted Healthfounding team · staff engineer · 2025–now
I designed and lead the agent platform that resolves medical-billing disputes for patients: long-running multi-agent workflows with explicit state, human-in-the-loop approvals, durable execution on Temporal, and evals. The product's AI-driven workflows run through it.
Multi-agent case automation · Real-time context engine · Production RAG over documents · Temporal durable execution · Evals & observability
Read the deep-dive →Work
Production AI since 2022, the last three years in healthcare: voice agents, RAG, multi-agent automation. Before that, a decade shipping full-stack products.
Assort Health
2023–24Founding engineer on a voice-AI contact-center platform for healthcare. Shipped intent routing that lifted automation from 70% to 85%, and ran infra for thousands of concurrent LLM patient calls.
ElevenLabs · OpenAI · Deepgram · Kubernetes
Blackbird.AI
2022–23Senior engineer on a disinformation-detection platform. Built the company's first LLM integration in under a month, demoed during the raise. Cut network-viz generation time by more than half.
Python · React · Elasticsearch · AWS · Celery
Sisense
2019–22Helped design and build the Python/R computing platform for the analytics product, and architected the RBAC and audit system for Periscope Data.
Ruby on Rails · React · Python
pMD
2016–18Full-stack and Android for a clinical charge-capture app serving thousands; built HL7 integrations.
Java · Android · MySQL · HL7
Side projects
DreamDraft
● liveAI children's-book creator. Fine-tunes Stable Diffusion on a family's own photos so the kid is the hero of the story.
The hard part was making personalized image generation reliable enough for non-technical parents.
Pixelbench
○ buildingEvaluation and monitoring for text-to-image models: measure quality, compare models, catch regressions.
Turning subjective image quality into numbers a team can act on is the whole problem.
Writing
From one prompt to a planner/executor split, a context registry, an event-sourced fact store, and evals that boot the real graph. Each piece exists because something broke without it.
Reading notes: what each paper changed in the agent platform, not a canon.
Contact
Open to staff and founding roles building agent systems.