Pranay Narula
Available · Summer 2026

Building AI systems designed for real-world conditions.

Prior work spans quantitative trading systems, LLM-backed enterprise tooling, and adversarial ML research briefed to U.S. Congress. Work centers on reliability under real-world conditions.

LLM Systems Adversarial ML AI Safety Quantitative Finance 0→1 Product UConn CS + Finance
Book a 30-min call → See my work ↓
Ask my portfolio
AI · knows my work
READY
Pranay's AI
Hi! I know Pranay's projects, research, and background. Ask me anything.
500+
Candidates evaluated in production
1st
Place, adversarial ML hackathon
3
Live production systems shipped
700K
Views on TechRoast founder feature
01 / work

Selected work — shipped, in production, used by real people.

A cross-section of internships and independent projects. References, scope, and outcomes available on request.

i.
2024 — 2025
Data Analyst · Research Engineer
Internship
RedVest TechCrunch-backed fintech

Proprietary trading algorithm and signal infrastructure for a Gen-Z investing platform.

Designed and shipped a proprietary trading algorithm powering the platform's portfolio allocation engine. Built the underlying research stack, owned model evaluation, and worked with the founding team to translate quantitative work into product decisions.

Scope
Algorithm + research stack
Stack
Python · PyTorch · quant libs
Stage
Production · live
portfolio.py● live
Portfolio Benchmark
signal_engine12m
ii.
Summer 2025
AI Engineer
Internship
STEPStone Enterprise HR tech

Custom hiring algorithm design system, powered by LLMs.

Designed and shipped a configurable LLM-backed hiring algorithm for structuring, weighting, and executing candidate evaluations. Demoed to two enterprise firms; evaluated over 500 real candidates in production.

Scale
500+ candidates
Reach
2 enterprise firms
Status
Under NDA
eval_pipeline.pyA-427
Technical
0.82
Role fit
0.68
Judgment
0.74
Comms
0.61
Ownership
0.79
Weighted score0.738
llm · rubric · configurable
iii.
Summer 2024
Founding Engineer
Internship
InterviewAI B2B SaaS · 0→1

Founding engineer on a B2B SaaS product — concept to paying customers.

Owned the entire engineering surface for InterviewAI's B2B launch: architecture, core product, payments, and vendor relationships. Shipped to live paying customers as founding engineer on the team.

Role
Founding engineer
Shipped
Full stack · payments
Outcome
Live · paying customers
launch.logv0
Spec · architectureshipped
Core product buildshipped
Payments integrationshipped
Vendor managementshipped
Public launchlive
v0
b2b · saaslive
02 / demos

Project demos — see it run.

Live recordings and interactive tools. More being added as projects ship.

Adversarial ML

PGD attacks on airport security classifiers.

Imperceptible image perturbations that flip classifier confidence on real security-context models. Hackathon 1st place.

Live demo
AI Research

AI bias finder — for LLMs.

Tool for detecting and surfacing bias patterns across large language models.

Live demo
Quant / Systems

End-to-end trading sim — zero-knowledge traders.

Full exchange stack: broker routing, orderbook matching engine with cancellation, and zero-knowledge trader patterns.Derived Markov processes from live order flow.

Live demo
03 / research

Research & independent work — where the interesting problems live.

Self-directed research and open-source projects alongside client work.

AI Safety · InterpretabilityCongressional briefing

Adversarial Latent Space Reasoning — decoding model internals under attack.

How LLM reasoning breaks down under adversarial pressure, and how to read it

Research into how adversarial conditions deform reasoning in an LLM's latent space. Presented to Congressman John Larson and congressional staffers via the Center for AI Policy.

research · ongoing
Computational BiologyOpen source

openMRNA — personalized mRNA sequence design.

Differential analysis of healthy vs. tumor sequences to generate candidate mRNA

Open-source pipeline for personalized mRNA candidate design — differential analysis to generate and rank neoantigen candidates. Experimental; research use only.

research · experimental
Robotics · MLActive research

LeRobot SO-100 — latent prediction for robotic control.

Predicting latent states for low-cost arm manipulation tasks

Research into latent space prediction models for the SO-100 robotic arm — using learned representations to improve control policy generalization on manipulation tasks.

research · active
Adversarial MLHackathon · 1st place

PGD attacks on airport security classifiers.

Projected gradient descent on image models deployed in security contexts

Adversarial attack study on image-classification models used in airport security workflows. Awarded first place; a practical argument for adversarial evaluation before deployment.

hackathon · 1st
04 / recognition
Congressional
U.S. Congress.

Briefed Congressman John Larson and staffers on adversarial latent space reasoning research via CAIP.

Institutional
OpenPhilanthropy

Supported institutional grant work as Treasurer of Beacon AI Safety at UConn.

Credentialed
Bloomberg BMC

Certified across Economic Indicators, Currencies, Fixed Income, and Equities.

Featured
TechRoast

Featured as a founder on a short that crossed 700K views.

05 / engagements

How I work — three shapes of engagement.

Dedicating this summer to a limited number of independent engagements — here's what that typically looks like.

01 · Direction

Technical direction & proof of concept.

For when you don't yet know what to build. I research the problem space, identify credible AI or quant angles, and run small experiments to find out what's actually worth pursuing — before any real commitment.

1–2 week sprint · fixed scope
02 · Build

Production AI prototype.

A working system — an internal tool, a customer feature, an evaluation pipeline — delivered fast enough to put in front of users, investors, or a board. Scope fixed before kickoff.

Typically 2–6 weeks · fixed price
03 · Audit

AI safety & red-team review.

Pre-deployment stress testing of AI features: what breaks it, what it says when it shouldn't, where it fails quietly. Delivered as a written report any stakeholder can act on.

Two-week audit · fixed scope
06 / reference
One of the rare young engineers who can move between deep AI research and practical product work without losing the thread. He ships, he questions, and he makes the team around him sharper.
Anthony P. Rotoli LinkedIn recommendation Read original →
07 / contact

Have an AI idea and a deadline? Let's talk.

The easiest way to find out whether the engagement is a fit is a 30-minute call. No pitch deck required — a paragraph on what you're trying to build is enough.

Book on Calendly → LinkedIn ↗
Emailpralo.nar@gmail.com
LinkedInpranay-narula
GitHubSslithercode
LocationStorrs, CT · Remote