{"$schema":"https://dhairya9.com/agent.json","name":"Dhairya Shah","role":"founding engineer — agentic systems","thesis":"Founding engineer scaling an early-stage healthcare-AI startup from zero to its first ten clinics, agents, infrastructure, and operator tooling, single-handedly. I reverse-engineer whatever has no API, and I build my own things on the side.","location":"Pune, IN (operates IST ⇄ PST)","tools":[{"name":"voice_agent","signature":"voice_agent(realtime=true, langs=[\"en\",\"zh\",\"es\"])","description":"production phone agents","detail":"Real-time phone agents on Vapi and Retell, live with real callers. Scheduling, screening questionnaires, multi-turn conversations that survive interruptions. Barge-in handling, mid-call tool calls, and handoff logic for the moment a human should take over. The multilingual fleet runs English, Mandarin, and Spanish with live language detection: one agent brain behind three phonetic front-ends."},{"name":"computer_use","signature":"computer_use(target=\"legacy_desktop_via_citrix\")","description":"UI automation where no API exists","detail":"Agents that operate legacy desktop software through the screen: reading the UI, clicking, typing, verifying. That includes apps running inside Citrix, where every frame is pixels and nothing is inspectable. Deterministic Playwright steps where possible, a vision agent loop (screenshot, decide, act via CDP) where the UI shifts. The rule: if a trained human can do it on screen, the agent can be taught to do it, with a human approving the result."},{"name":"extract","signature":"extract(engine=\"ocr+llm\")","description":"documents → structured data","detail":"Pipelines that turn messy scans and faxes into clean structured records: PaddleOCR for the pixels, Claude for the judgment, evals for the honesty. Field-level accuracy is measured against labeled sets, not vibes. Running at production scale against the documents nobody designs for: skewed scans, handwriting in the margins, forms from regulated workflows that predate the web."},{"name":"sre_agent","signature":"sre_agent(mode=\"read_only_default\", confirm=true)","description":"investigate → heal → draft postmortem","detail":"An internal SRE agent with scoped access to GCP (logging, Cloud Run, Cloud Tasks), GitHub, the voice provider, and the database through a read-only proxy. It detects anomalies, investigates the likely cause with a confidence score, and proposes self-heals for known failure modes: a stuck-state reaper, releasing stale held slots, unsticking a referral. Safety posture is the product: read-only by default, no production mutation without an explicit human confirmation gate, and auto-drafted postmortems that become reusable playbooks."},{"name":"build","signature":"build(stack=[\"rust\",\"tauri\",\"react/ts\",\"fastapi\",\"gcp\"])","description":"the cockpits the agents live in","detail":"Agents need a control surface humans trust. Mine is a Tauri 2.0 desktop app (Rust core, React/TypeScript front) where operators review, approve, and override what agents did. Overlay UIs inject agent output directly into third-party software. Behind it: FastAPI + SQLAlchemy on GCP, queue-driven state machines, and EDI/X12 protocol integrations, the unglamorous plumbing that decides whether an agent system is a product or a demo."}],"runs":[{"id":"run: voice_fleet","status":"completed","objective":"Answer and place phone calls no one has time for.","outcome":"Live with real callers daily; call outcomes classified and audited by a second model.","links":[]},{"id":"run: citrix_operator","status":"completed","objective":"Operate desktop software that has no API, through Citrix.","outcome":"Replaced hours of daily manual data entry; runs unattended with operator review.","links":[]},{"id":"run: document_pipeline","status":"completed","objective":"Turn inbound fax queues into structured data.","outcome":"Production scale; measured field-level accuracy in the mid-to-high 90s on labeled evals.","links":[]},{"id":"run: sre_agent","status":"completed","objective":"Detect, investigate, and safely heal agent-runtime incidents.","outcome":"Read-only by default; no prod mutation without explicit human confirm.","links":[]},{"id":"run: healthadmin_models","status":"in_progress","objective":"Explore specialized agentic models for long-horizon healthcare-admin workflows.","outcome":"Early stage on purpose: data and benchmark pipelines first, models next. Reproductions before claims.","links":["https://arxiv.org/abs/2604.09937"]},{"id":"run: physical_ai","status":"in_progress","objective":"Learn physical AI by building, not reading.","outcome":"Just started. The point is reps and taste, not a launch.","links":[]},{"id":"run: schedlab_rs","status":"completed","objective":"Find out when shortest-job-first actually wins in an LLM inference scheduler, and whether that answer survives your cost-model assumptions.","outcome":"Predicted-SRPT captures ~97% of oracle goodput (regret 0.012 vs FCFS 0.358), but the winning policy is cost-model-dependent in 16 of 45 regimes. A single-equation sim would report that as fact. Findings replicate on a real Azure LLM trace (~98% of oracle). MIT.","links":["https://github.com/DhairyaShah981/schedlab-rs","/logs/when-does-sjf-stop-winning"]},{"id":"run: voice_eval_harness","status":"completed","objective":"Make voice agents testable before they embarrass you.","outcome":"Caught 13 fatal config bugs across 8 production agents on day one. Apache-2.0.","links":["https://github.com/DhairyaShah981/voice-eval-harness","https://voice-eval-harness.vercel.app"]},{"id":"run: skill_pack","status":"completed","objective":"Package the agent-engineering playbook as a skill pack.","outcome":"MIT, v0.4.x; iterating in public.","links":["https://github.com/DhairyaShah981/healthcare-skills"]},{"id":"run: mcp_server","status":"completed","objective":"Ship an MCP server so agents can consume clinical tooling safely.","outcome":"Open source; audit-log ↔ LLM-trace correlation included.","links":["https://github.com/DhairyaShah981/fhir-mcp","https://dhairya-fhir-mcp.fly.dev"]},{"id":"run: research","status":"completed","objective":"Publish ML research that survives review.","outcome":"Published, ACM JCSS 2025 (DOI 10.1145/3748334).","links":["https://scholar.google.com/citations?user=O8YcmX8AAAAJ&hl=en","https://doi.org/10.1145/3748334"]}],"checkpoints":[{"id":"ckpt: founding_engineer","period":"Feb 2026 to now","title":"Founding engineer · first India hire","org":"TriFetch · Nexus Venture Partners · $1.9M pre-seed"},{"id":"ckpt: bny","period":"Jul 2024 to Jan 2026","title":"Software Development Engineer 1","org":"PershingX, BNY (Wove platform)"},{"id":"ckpt: first_run","period":"Jun to Jul 2023","title":"Software Development Intern","org":"BNY"},{"id":"ckpt: iitgn","period":"2020 to 2024","title":"B.Tech, Computer Science and Engineering","org":"IIT Gandhinagar · CPI 8.57/10"}],"actions":[{"name":"human_handoff","description":"Contact the human. The only exposed action.","email":"dhairyarup@gmail.com","github":"https://github.com/DhairyaShah981","linkedin":"https://www.linkedin.com/in/dhairyashah0819/","twitter":"https://x.com/dhairya_dhiru99","scholar":"https://scholar.google.com/citations?user=O8YcmX8AAAAJ&hl=en"}],"telemetry":null,"note":"This file is served for your agent's convenience. My people will talk to your people's agents."}