i own enterprise deployments of real-time voice and sms agents end to end: agent design, crm integrations, typescript backend features, and the observability and incident systems behind them. i sit between customers and core engineering, turning field failures into scoped tickets, design docs and production changes.
by the numbers
scheduling failure codes19a structured failure taxonomy i built, so every failed booking carries a reason you can query, alert on and fix.
crm & scheduling systems integrated4+
servicetitan
housecall pro
pestpac
fieldroutes
+ others
via apis, adapters and webhooks, with platform features that keep agent behaviour consistent across crms.
a call, end to end
customercall or sms→voice agentvapi · elevenlabs→prompts & blueprintsknowledge-base variables→tool callsbooking logic→crm writeservicetitan · housecall pro · …
enterprise ownership. own multiple enterprise accounts end to end, from agent build and crm integration to launch and production reliability.
llm agent engineering. configure and deploy production llm agents: prompt and blueprint composition, tool wiring, knowledge-base variables and booking logic across the voice platform layer.
integrations & platform. integrate agents with crm and scheduling systems, and design platform features that cut manual onboarding work.
production debugging. debug distributed failures spanning telephony, the llm, tool calls and crm writes using logs, traces, sql and code-path analysis, and attribute regressions to the exact config change or commit.
backend engineering. ship features and fixes in a typescript monorepo (next.js, postgres, inngest), covered by tests and type checks and merged through pr review.
founding ai engineer · sally, conversational & real-time voice ai
i architected and shipped sally, a full-stack ai sales agent that talks to leads in real time for 30+ u.s. enterprise clients, and the guardrails that keep it honest.
by the numbers
conversations a month50K+real-time sales conversations handled by sally across clients.
u.s. enterprise clients30+
lift in lead-to-close conversion38%
before100
with sally138
conversion indexed to 100 before sally.
fewer hallucinations60%
before100
after40
guardrail layers, structured output validation and automated regression testing.
speech-to-speech latency500–800ms
050010001500ms
from the moment a caller stops talking to the moment sally starts, with streaming stt/tts and neural vad turn-taking.
delivered enterprise automation on custom mcp servers with multi-model orchestration (gpt-4, claude, gemini) on fastapi and kubernetes, automating lead processing and crm enrichment to lift client arr/mrr.
what i did
architected and shipped sally, a full-stack ai sales agent: llm orchestration with prompt chaining, tool-use routing, context windowing and eval pipelines.
real-time voice with streaming stt/tts and neural vad turn-taking at 500–800ms speech-to-speech latency, driving a 38% lift in lead-to-close conversion.
hallucination-reduction framework: guardrail layers, structured output validation and automated regression testing, cutting hallucination rates by 60%.
infrastructure on aws with distributed tracing, ci/cd and docker/kubernetes.
computer vision that clinicians actually used: diabetic retinopathy screening and cbct-based dental implant planning, built and clinically deployed across 30+ clinics.
by the numbers
screening accuracy99%+
diabetic retinopathy screening.
retinal images trained on50K+with gan augmentation to widen the training set.
persistent semantic memory layer for autonomous ai agents
an open-source, mcp-native memory server. agents remember across sessions, link what they learn into a knowledge graph, and consolidate experiences into lasting memory.
by the numbers
search latency<5mshybrid retrieval over the whole memory store.
steps in memory consolidation7
1234567
episodic memories consolidate into lasting knowledge.
b.e. computer science · sir m. visvesvaraya institute of technology, bangalore · 2022 – 2026
i had a good college life: friends, the basketball team, late nights that had nothing to do with exams. but most of my time didn't go into studying for grades. it went into building things real people used.
i believed then, and still do, that you learn more from shipping software and watching it fail than from any textbook. so i spent college building startups and real-time systems, and spending weekends at hackathons.
four years, side by side
college
medai
100x
avoca ai
202220232024202520262027
half my degree overlapped with running a company or building one.
startups
building, not memorising
i founded medai while in college and took clinical computer vision into 30+ clinics. in my final year i joined 100x as a founding ai engineer and shipped a voice agent handling 50k+ conversations a month. both taught me more than any semester did.
won several
hackathons
a lot of my weekends went to hackathons, and i won multiple of them. win or lose, every one forced me to ship something real in a weekend, which is a skill no lecture teaches.
philosophy
learning through failure
grades measure what you remember. building measures what you can do. i'd rather break a production system at 2am and learn exactly why than ace a test on it.
campus
outside the classroom
played basketball for the college team and started karuna, a drive to get puppies vaccinated.
no papers yet. these are the questions i keep coming back to. if you're working on them, i'd love to talk.
interest
ai safety
i ship agents that act on their own in front of real customers. that makes safety a practical question for me, not an abstract one: how do you know what an agent will do, and how do you keep it that way?
how do you evaluate an agent whose behaviour depends on what it remembers?
can an agent's memory be audited, corrected and made to forget?
when should an autonomous system stop and hand control back to a human?
interest
consciousness through memory
cortex started as infrastructure: give agents memory that persists. but persistent memory raises a bigger question. if a system carries its past forward, reflects on it and changes because of it, what is it becoming? i think deep memory architectures are one of the more concrete paths toward something like machine consciousness.
what is the smallest memory architecture that gives an agent a sense of continuity over time?
how should episodic memories consolidate into semantic knowledge, the way they do in us?
can an agent build a model of itself from its own history?
what i do when i'm not deploying things. (sometimes still deploying things.)
obsession
building startups
my favourite hobby is also my work. i love building things at scale that help a lot of people, whether that's screening in 30+ clinics or 50k+ conversations a month.
founded
karuna
karuna is a drive i started to get puppies vaccinated, so they survive their first months without catching parvo, lepto and other diseases that kill so many of them early.
🐾🐾🐾🐾🐾
court
basketball
played for my college team. (tap it)
underwater
aquascaping
building nature underwater. what fascinates me is the science: managing nitrates, the nitrogen cycle and co2 so plants grow, and watching a glass box develop into a mature ecosystem of its own.
fish waste→ammonia→nitrite→nitrate→plants
daily
training
working out and staying athletic. it keeps the rest of this running.