Healthcare phone lines, where the caller is a patient and a wrong answer actually matters. At VoxyHealth I sit forward with the customer: pull their real call recordings, catalogue what's on their lines, then build the agents and iterate on live traffic until the numbers prove it.
Lia is a voice AI agent Auskin built. Conversations may be recorded.
What a good call sounds like.
The shape of the work: a live phone line, a real scheduling system behind it, and a caller who gets what they came for in one conversation. Scroll through a call.
Illustrative transcript. No real patient, clinic, or call data.
I started in sales development, moved to customer success owning two AI products, then into product to build the agent side itself. Selling AI, implementing AI, building AI. That arc is why my agents work: I've sat on every side of the table where they land.
I start inside the customer's world: their ops team, their call recordings, their real day. What actually happens on the line, not what the spec says happens.
3,298 real calls classified in three days into 54 call types, which became roughly 40 flows in the live build
Scenario maps, transfer rules, tool contracts, prompt architecture, and the success criteria the release will be judged on. The agent exists on paper, agreed with the customer, before it says a word.
911 and red-flag safety routing from day one
Eval-gated releases, multi-LLM A/B tests, and pilot calls I place myself against the real integration before anything goes live.
about one new agent a week stood up, sustained March through July
Live traffic is the real eval. I read the transcripts and sort every gap into one of four buckets: prompt issue, design issue, platform debt, or something that shouldn't be automated at all. Each bucket gets a different fix, and fixing the wrong bucket is how teams burn weeks.
a daily ground-truth pipeline grades every claims call; 268 hand-labelled for one ROI report
The meta-skill underneath the method: I find the gaps in my own understanding fast, and close them faster.
Before a build or a big call, I map what I don't know and run deep research on it, then turn it into audio briefings that run in the background of my day. Ground truth first, opinions second.
Ideas become working tools the same day: internal utilities, prototypes, this site, and Lia, the agent living in the orb up top. I set that tooling up for my whole prompting team.
Off the clock it's physics and quantum computation: logical qubits, verifiable quantum advantage, the roadmaps under the headlines. Curiosity, not expertise. But every frontier eventually needs people who can make it useful inside a real organization, and I want a head start on whichever one comes next.
I built the agent that answers their front desk, wired into their EHR and live since August 2026. One phone number covers every office, and callers name a city, a doctor, a body part, or nothing at all, so placing the caller is the first job on every call. I made the handoff a contract, not an association: every transfer is keyed to the exact phrase the agent speaks, because a natural-sounding synonym dials the wrong desk. Most of them run warm, so an outbound leg briefs the human first and bridges the caller only on a yes.
My deepest payer relationship: one California health plan, five agent lines, built and iterated as one fleet. Rebuilding the Medicare Advantage claims line around call outcomes lifted fully-AI-handled containment from the 10-20% range to 60-70% on best cuts, as of June 2026. I ran five LLMs live against a single rubric, because a model earns its slot on data, not on feel. What started as one claims line is now a five-line family.
Outbound agents that coordinate patients, pharmacies, and providers across multiple calls to close screening and medication gaps. A 3-layer prompt architecture runs five care gaps per patient under a single master agent, with clean AI-to-human coordinator handoffs. One trade-off I'm proud of: a wrong automated answer on medication checks costs patient trust, so I removed the AI-to-pharmacy calls in favor of a human check. Some loops belong with humans.
Moved about nine production agents to a small, fast model in four weeks. The unlock is prompt architecture: a lean base prompt plus one card per scenario, with a single trailing guardrails block. The token drop is what buys a live phone call its latency.
From the first client workshop to a live pilot, and past it. I embed with the customer's team, scope the engagement, and set what success means before the build starts. The solution takes whatever shape the problem needs: a voice agent, a platform automation, an API integration. Then I read how it performs live, diagnose the gaps, and prove the outcome.
I own the agent roadmap: what we build, what good means, and the rubric that proves it. Success criteria are agreed before the build starts, and eval pass-rate is the release gate.
30+ agents built on ElevenLabs, 20 live in production. Prompting architecture, model routing, latency budgets, bilingual support, and speech discipline for small models.
SDR, then CSM, then PM. I've been the person on the call at every stage, and I still run the customer meetings and business reviews for my agents.
I hand engineering exact contracts: tool schemas, validation specs, release plans. I review PRs for agent-behavior impact, and live call failures come back as clean, testable asks.
I ship internal tools, transforms, and prototypes with Claude Code, and lately two public platform builds outside voice: a Sheet-to-PSA time-entry sync on Workato, and a form that writes its own requirements doc. Enough to build, debug, and partner credibly with engineering, not to replace it.
VoxyHealth
I own the prompting and evaluation side of healthcare voice agents end to end: discovery, prompt architecture, eval rubrics, phased go-lives, live pilot calls, and the specs engineering builds against. I've built 30+ agents on ElevenLabs across 6 enterprise clients, 20 of them live in production running about 3,500 calls a day as a fleet, and I mentor the team's two prompting interns.
Klenty
Owned customer success for two AI products: SchedX, an inbound AI voice platform, and SDRx, an AI SDR for email, LinkedIn, and calling. Onboarded and implemented clients across four verticals and four CRMs (HubSpot, Salesforce, Zoho, Pipedrive). The main bridge from customers back to product and engineering.
Klenty
B2B SaaS sales, starting in my final year of college. Generated 15+ qualified leads and booked 8+ meetings a week, consistently above target. The customer fluency behind how I scope and demo agents started here.
Aptitude Guru Hem
Led a small team to ship the company's learning management system and marketing website end to end, working week to week with the founders.
Loyola-ICAM College of Engineering and Technology, Chennai. B.E. Computer Science, 2020 to 2024, 8.63 CGPA.
My first real software shipped here: a ten-month campus internship building the college's student portal, with a letter of appreciation at the end of it. Build a thing, watch real people use it, fix what they trip on. I've been chasing that loop ever since.
College is also where the pattern set. I'm the one who organises the fest, hypes the room, and then goes quiet for a week to lock in and build the thing. Both gears, on purpose.
The receipts are still public: the student portal, the workshop repos, all of it. Old repos on GitHub →The flagship writeup: how I build production voice agents, from prompting architecture to model choice for low-latency telephony, with six worked samples for ElevenLabs. Around it sits a family of builds in a different shape entirely, platform automations and API integrations, because the discipline travels: define the contract, prove it on the live system, design the failure path first.
Google Sheet to PSA time entries through a daily Workato recipe. Idempotent, with the error queue living in the sheet where ops already works.
A form submission becomes a requirements document, fully native, no middleware. The load-bearing step wasn't documented anywhere; I verified it by building it.
The API sandbox both builds grew out of: pagination, rate limits, and the contract traps the docs don't warn you about.
auskinimmanuel@gmail.com · +91 90035 43634
Based in Chennai. My customer calls already run on US hours.