Hritvik Gupta hritvikgupta

Hi, I'm Hritvik — an AI-native engineer

I build agents that do real work — not demos.

Multi-agent orchestration · longitudinal memory · RAG & retrieval · eval harnesses · LLM infrastructure

About me

I'm Hritvik — an AI engineer at Penn Medicine (Verma Lab), where I ship voice and chat systems that real patients actually use. Outside of work I build autonomous agents: things that read your code, operate your cloud, read the literature, and hand back a result you can verify.

My bet is simple — the interesting part of an agent isn't the model, it's the loop around it: memory, tools, and evals. Most agents feel dumb because they forget, not because they can't reason.

name:         Hritvik Gupta
role:         AI Engineer @ Penn Medicine · Verma Lab
education:    MS Computer Engineering, UC Riverside
thesis:       "agents should do real work — not demos"
building:     [autonomous agents, LLM infrastructure, agent evals & benchmarks]
languages:    [Python, TypeScript, Swift, SQL, C++]
now:          multi-agent orchestration · longitudinal memory · self-hosted agent infra
ask-me-about: [RAG at scale, agent evals, multilingual speech pipelines, genomics ML]

100K+
_{patients served by the voice & chat system I built}

27M
_{SNPs processed on Argonne's Aurora supercomputer}

10M+
_{multilingual research documents in NLP pipelines}

4
_{peer-reviewed publications}

What I'm building

More projects →

Project	What it is
chytra	AI-powered research and creative canvas — a Figma-style design surface wired to 200+ models, with a graph-memory architecture that connects ideas, documents, and findings
worklone	Next-generation AI spreadsheet and agentic framework — natural-language data workflows, a built-in Data Scientist agent, and multi-agent request routing
probeqa	Agentic QA — a real testing agent that drives the app instead of asserting on mocks
voiceai	Real-time voice agent stack: streaming STT → LLM brain → TTS
docuwriters	Documentation that writes and maintains itself from the codebase

How I build agents

flowchart LR
    U(["User · event · schedule"]) --> P
    P["Perception<br/>speech · code · logs · papers"] --> R
    R{{"Reasoning<br/>plan · route · decompose"}} -->|delegate| S["Specialist sub-agents<br/>research · rank · reproduce"]
    R -->|tools| A["Action<br/>PRs · queries · deploys"]
    S --> A
    A --> V["Verification<br/>evals · deterministic checks"]
    V -->|regress| R
    V -->|ship| O(["Real-world side effect"])
    M[("Longitudinal<br/>memory")] <--> R

    classDef n fill:#F6F8FA,stroke:#C6CDD5,color:#16191D
    classDef h fill:#FDF0EC,stroke:#E14D2A,color:#16191D
    class U,P,S,A,V,O n
    class R,M h

Tools over talk. An agent's output is a side effect in the real world — a merged PR, a rolled-back deploy, an escalation — not a paragraph that reads well.

Memory is the hard part. Continuity across sessions beats brilliance inside one. Every agent I ship gets a working-memory document it maintains itself.

Evals or it didn't happen. A benchmark harness with deterministic checks goes in before the agent meets a user, not after it embarrasses one.

Self-hostable by default. Your data, your infrastructure, your keys. Anything holding production credentials should run where you can watch it.

Experience

gantt
    title       Career timeline
    dateFormat  YYYY-MM
    axisFormat  %Y
    todayMarker off

    section Industry
    Data Analyst · Cognizant               :c1, 2021-08, 2022-08
    AI Engineer · Penn Medicine            :active, p1, 2024-07, 2026-09

    section Research
    Graduate Researcher NLP · UC Riverside :r1, 2022-10, 2023-12

    section Education
    MS Computer Engineering · UCR          :e1, 2022-09, 2023-12

AI Engineer · Penn Medicine · Jul 2024 – Present

Built an AI voice & chat system for Perception Care used by 100K+ West Coast patients — a multilingual speech → RAG → LLM pipeline on LlamaIndex, FAISS, LangChain, FastAPI, and Docker.
Engineered data generation for speech and language model retrieval, vector search, and prompt-routing workflows — ~28% lower response latency, +22% clinical-text retrieval accuracy.
Enhanced the PLATLAS genomics platform with ML-based variant ranking and phenotype-similarity scoring; ran 27M-SNP Nextflow pipelines on Argonne's Aurora supercomputer.
Developed PySpark + Delta Lake pipelines standardizing 30+ clinical datasets into OMOP, enabling real-time cohort building and disease-trend dashboards.

Graduate Researcher (NLP) · University of California, Riverside · Oct 2022 – Dec 2023

Built large-scale NLP pipelines (Python, Spark, SQL) over 10M+ multilingual research documents, improving tokenization and embedding generation speed by ~40%.
Optimized RAG systems with LlamaIndex + LangChain, raising scientific-text retrieval accuracy by 22%.

Data Analyst · Cognizant · Aug 2021 – Aug 2022

Built Python + SQL ETL pipelines over 20K+ HR, payroll, and marketing records across 50 datasets, improving data accuracy by ~35%.
Developed Scikit-learn + SAS predictive models for attrition and hiring demand, improving workforce planning for 5,000+ employees.

Tech stack

Languages
AI & agents
Backend
Data
Cloud & infra

Stats

Publications & recognition

Year	Work	Venue
2025	Levin, M.G., et al. (incl. Gupta, H.) — Genome-Wide Assessment of Pleiotropy Across >1000 Traits from Global Biobanks	medRxiv
2021	Gupta, H. & Patel, M. — Text Summarization: LSA Topic Modelling with BERT	AI Smart Systems
2021	Gupta, S. & Kal, H. — Microstate EEG Analysis via RNN	i-PACT
2020	Patel, M. & Gupta, H. — Extractive Text Summarization Using ELMo	IEEE I-SMAC

Health-Tech Innovation Accelerator Award — Penn Health-Tech, 2025 CIRCA: Voice-AI for general healthcare services to patients.

Let's build something

I'm always up for a conversation about agents that have to work in the real world — production credentials, messy data, and users who notice when it's wrong.

_{~/hritvik $ agents --status → shipping}

Navigation Menu