AI Researcher & Engineer / UC San Diego At ServiceNow now · looking for full-time roles

PRANAV SOMA.

I build AI & ship it.

I cut AI Voice Agent latency 45% at ServiceNow and shipped the platform to production, validated live with a real enterprise customer. At UCSD I lead a campus-wide agentic platform and an AI tutor running for 8,000 students. Bachelor's in two years, master's in just over one at a 4.00 GPA. On the side this fall, I start building the software an autonomy stack runs on.

See my work Résumé Email me
Pranav Soma
Pranav SomaSan Diego · open to the Bay
0Master's GPA
0students on my AI tutor
0voice-agent latency cut
0interns; I ranked first
0on the platform I lead
PythonPyTorchLLM fine-tuningLoRARLHFGraphRAGMCP serversReactFastAPIAWSXGBoostOpenCVDockerMongoDBC++TypeScriptNodeSensor fusionDistributed systemsROSTransformers PythonPyTorchLLM fine-tuningLoRARLHFGraphRAGMCP serversReactFastAPIAWSXGBoostOpenCVDockerMongoDBC++TypeScriptNodeSensor fusionDistributed systemsROSTransformers

A demo is a guess. Production is the proof.

Ask me for a link. I'll usually have one.

Experience

Where I've worked

Five teams, five problems that looked nothing alike. Four of them I shipped. The fifth starts in September.

Starting Sep '26Incoming · part-time

Founding Software Engineer @ a startup still in stealth

A few hours a week alongside my master's, because the problem is too good to pass up. In September I start building the core operating system an autonomy stack runs on, from nothing: hardware-agnostic, still standing when a piece of it falls over, able to scale out across a lot of distributed nodes. On top of that sits the autonomy work (multiple agents coordinating, planning trajectories, splitting resources between them), a perception stack that fuses many sensor feeds into one read on what is actually happening, and the low-latency dashboards that hand an operator a single picture of all of it. It has to hold up when the network is bad, and it has to run on real hardware, not just in sim. This is the one thing on this page I haven't shipped yet. Ask me in a year.

AutonomyMulti-agent coordinationSensor fusionC2 / telemetryDistributed systemsHardware-in-the-loop
Summers '25 & '26Live with customers

Software Engineer Intern @ ServiceNow

I took the AI Voice Agents platform off a version of AWS Lex that was about to be killed, moved it to the new one ahead of schedule, and got it into production. We validated it live with an enterprise customer before my internship was even over, which unblocked the deployments waiting behind it. I also rebuilt the language-understanding pipeline across 8 intents and cut AI Voice Agent latency by 45%. They gave me a return offer and the Technical Skills Award (top intern of 800+ at UCSD's symposium), and I'm back interning there again this summer.

AWS LexLambdaJavaScriptPythonPostgreSQL
2025 - nowTeradata-funded

AI Research Fellow & Agentic Systems Lead @ Lab for Emerging Intelligence

I lead an agentic AI platform that 7,000+ students, faculty, and staff at UCSD use. That means custom MCP servers over the university's APIs and a workflow engine I built to stay reliable when things go wrong. I also built and shipped an AI tutor that's now running for 8,000 students across UCSD, SDSU, and Cal Poly Pomona. We A/B tested it: TA hours dropped 64% and engagement went up 87%.

MCP serversLoRA / SFT / RLHFFastAPIMongoDB
2024 - 2025Health AI

Graduate Researcher @ Qualcomm Institute (HXI Lab)

I worked on LifeSaver, an assistant that reads sensor, audio, and video data and turns it into something a health database can answer, built for emergency services, the elderly, and rural care. I trained multimodal models on the San Diego Supercomputer Center and used GraphRAG over the UMLS medical ontology to keep answers accurate and fast enough for emergencies. I'm first author on the lab's poster.

Multimodal LLMsQwenGraphRAGUMLS
2024Intern lead

Software Engineer Intern @ Work4Flow

I led a team of 6 interns. We shipped GenAI Observer, a tool that scores the prompts engineers write and tells them how to make them better, plus an integration connecting ServiceNow and Smartsheet with some LLM categorization. Both got the green light from the PM to go to customers.

LLM evalServiceNowSmartsheet
Research

What I'm researching

Four problems I haven't been able to put down. Each one is interactive below, go ahead and poke at it.

01

Computational redistricting: fairness in congressional maps

★ UCSD CSE honors thesis & master's thesis

My honors thesis, and now my master's thesis too. It draws and scores U.S. congressional maps over Census data, anywhere from 158K to 5.5 million blocks per state. I built three different map generators (a greedy heuristic, a Monte Carlo sampler, and an XGBoost model with graph partitioning) and wrote the fairness metrics that grade them, things like the efficiency gap and Voting Rights Act compliance. It reproduces the real district counts for Alabama, California, and Texas.

8 districts · click the map to redraw
XGBoostpymetisMonte CarloGeoPandasCensus data
02

Agentic systems for the whole university

Lab for Emerging Intelligence · Teradata-funded

I lead the Agentic Systems project at LEI (under Dr. Ramamohan Paturi, Matthew Clegg, and Dr. Elizabeth Lyons): enterprise, workflow-driven agentic AI for 7,000+ UCSD staff, faculty, and students. Custom MCP servers over the university's APIs, research into automated skill-generation and optimization, and a novel workflow executor I built to stay intelligent, deterministic, and graceful when a task goes sideways.

MCP serversWorkflow executorSkill optimizationFastAPI
03

Smart Learning Hub & a deployed AI tutor

Lab for Emerging Intelligence · LLM reasoning & fine-tuning

A deployed AI tutor and the platform behind it. To keep it reliable at scale, I built a two-layer evaluation pipeline: one layer monitors tutor outputs live; an LLM panel-of-judges above it composes and stores each student's learning and personality metrics so the teaching adapts. Running for 8,000 students across UCSD, SDSU, and Cal Poly Pomona, with 64% fewer TA hours and 87% higher engagement (A/B tested).

LLM reasoningLoRA / SFT / RLHFLLM panel-of-judges evalsLive output monitoringBayesian knowledge tracingLearner & personality modelingPython
04

AILA: an AI living assistant for emergency & rural care

LifeSaver · Qualcomm Institute / Calit2 · HXI Lab

An assistant that turns messy real-world signals (sensors, audio, video) into structured answers a clinician or first responder can use. I trained the models on a supercomputer and grounded them with GraphRAG over a medical ontology so the answers are both right and fast enough to matter in an emergency. I also wired up the robotics side that drives the actual hardware.

Multimodal LLMsQwen SFTGraphRAGUMLSSupercomputer

I'll take the harder problem if it's the one that actually helps someone.

The elderly, an ER, students without enough TAs. I keep ending up there on purpose.

Projects

Stuff I've built

The things I build when nobody's grading them. Keep scrolling, this one goes sideways.

♪ sheet music ↔ audio

ORCA Best Project

SDTC, picked 1 of 30 from 200+ · SAIRS 2025

ORCA (Orchestration and Recognition for Composition and Arrangement) is an AI music tool that turns sheet music into audio and audio back into a score. OpenCV reads the page, neural nets split the instruments, and a CNN writes the score. 98% accuracy, verified four ways including by real musicians.

OpenCVCNNFlaskDockerReact

Spay.LA

Engineering manager, team of 11

A fundraising and advocacy platform I led for a real client. It raised $75K. I ran the team and owned everything from scoping with the client to shipping.

spay.la →

AIDE

Innovate research group

AIDE (AI Integrated Developer Environment): an intent-based, zero-touch IDE that turns what you describe into working code.

AgentSpace

Innovate research group

A two-way marketplace: pros pour their personality and toolset into a hireable agent. Two-way GraphRAG, MCP + A2A.

BioVeritas

AI for scientific rigor

A tool that checks how sound and honest a piece of biology research is, using retrieval to back up what it flags.

Union Station Housing

Software developer

A listing platform for one of greater LA's largest housing providers. Photo and video upload, form validation, and bulk data exports for their team.

Skin Lesion Detector

Innovate research group

An image classifier for spotting skin lesions, built to run on small on-device hardware.

Virtual Trial Room

Innovate research group

A 3D fitting room using deep learning and computer vision so you can see how clothes look without trying them on.

Leadership

What I've started

I tend to start things. These are the communities I've built and helped run.

Real World AI Network

Co-founder & president

I co-founded a 7,000+ student group that spans UCSD's engineering, data science, and business schools. It exists to push AI research, building, and startups across the whole university.

rain.ucsd.edu →

University AI Alliance Nation's largest

Founding board, VP of external affairs

I helped start the biggest student-led AI initiative in the country, with people from MIT, Stanford, Caltech, and Cornell. I run the outside relationships: VCs like a16z speedrun and Techstars, and folks from Google and Meta.

CSE Society @ UCSD

President

I lead UCSD's oldest computing org: 200+ members across 21+ projects. We run career fairs, conferences, talks, and a research journal, and we partner with the Linux Foundation and Cisco on open source.

csesucsd.com →

Innovate Research Group

Founder

I started a research group that brings AI into other fields, working with UCLA, UCSB, and UCSC. I ran 30+ people across ORCA and a couple of other projects.

Next, I want to do this where the models are biggest and the stakes are real.

The hardest version of my problem lives at the cutting edge. Wherever that is, that's where I want to be.

Recognition

Some things I'm proud of

2026
YC Startup School 2026
Y Combinator's Startup School, Chase Center SF.
2025
ServiceNow Technical Skills Award TOP INTERN
Top intern out of 800+ at UCSD's Summer Internship Symposium.
2025
Cum Laude & CSE Department Honors
UC San Diego, B.S. Computer Science. Provost Honors every quarter.
2025
Best Project, SDTC
ORCA, chosen as 1 of 30 out of 200+ applications. Also shown at SAIRS 2025.
2022
National Merit Finalist
Also AP Scholar with Distinction and a National Speech & Debate award.
2022
USACO Silver
USA Computing Olympiad.
About

I grew up building robots and somewhere along the way got obsessed with putting AI in front of real people.

I'm from Cupertino, I move fast (undergrad in two years, master's in a year and a quarter), and I run pretty wide: systems, ML, theory, full-stack, robotics. The constant is that I like taking something that only works in a notebook and getting it all the way to someone who needs it, a student short on TAs, an ER, a family in a town with no doctor nearby.

The range isn't an accident. I came up through competitive robotics and debate and a USACO Silver, where you only get credit for the thing that actually works on the day. I'd rather be the person who can carry a problem from the math to the deploy button than stop at either end.

The other half of my time goes into building rooms for people: RAIN, the University AI Alliance, CSE Society. The crowd around a problem ends up mattering as much as the problem. Long term I want to keep doing both, the research and the rooms it happens in.

Contact

Let's talk.

Résumé LinkedIn GitHub

Targeted résumés: Software, Machine Learning, Forward Deployed

Actively looking for full-time roles. Based in San Diego, open to relocating anywhere in the SF Bay Area.