UCSC AIEA Lab — Prof. Leilani Gilpin USACO Platinum — Top Division Founder, ThePharmaBridge

Siddharth Shukla

Building the occupational health layer
for U.S. firefighters.
Proving LLM safety tools fail at the structural level.

View Research ThePharmaBridge ↗
0 LLM concepts
audited
0 solo research
papers
0 USACO Bronze
competitors ranked
0 Solvathon entries
beaten (1st place)
scroll
Research

Auditing whether safety interventions for LLMs actually function the way researchers claim. Consistent finding: they don't.

EMNLP ARR ’25 ICML Mech-Interp ’26 ICML Hypothesis Testing ’26

Structural Specificity Testing (SST)

A statistical auditing framework proving behavioral steering success does not equal structural concept recovery. Large-scale causal audit across 950 concepts using SAE decompositions on Gemma-2-2B-IT. Built CryptoBench — 450 security-domain concepts. Core finding: the most behaviorally effective steering vectors are the least structurally specific.

PyTorch · SAE-Lens · TransformerLens · Lambda A100 Solo first author
ICML AI for Good ’26

CIPHER

Benchmark measuring cryptographic vulnerability incidence in LLM-generated Python code. Insecure / neutral / secure prompt variants per task, cryptography-specific taxonomy, line-level attribution. Finds explicit secure prompting reduces targeted issues but does not eliminate vulnerabilities overall.

Python · OpenRouter API · Automated Scoring Solo first author
Independent Research

Latent Adversarial Concept Steering

Bypasses the Dead ReLU problem in Sparse Autoencoders using a custom Softplus Straight-Through Estimator. Successfully spoofed Anthropic-style SAE monitors, forcing Llama-3 and Gemma-2 to output hazardous content while blinding detection features. Fundamental architectural vulnerability.

PyTorch · llm-attacks · bfloat16 precision Solo author
B2B Deliverable — UnitZero

RoboEval

Bypassed physics simulators as an evaluation bottleneck entirely. Rebuilt from first principles: offline kinematic benchmark via signal processing — Savitzky-Golay filters, Welch’s Method, PSD. Containerized via Docker + AWS Batch. Enables 10TB dataset processing at scale.

Python · SciPy · Docker · AWS Batch Sole architect
Work

ThePharmaBridge

California Nonprofit · Incorporated Feb 2026
thepharmabridge.org ↗

9% higher cancer incidence. Firefighters. No infrastructure to act on it.

A health-tech platform that ingests occupational exposure data (PFAS, diesel, WUI smoke) and uses a custom RAG pipeline to generate clinical screening PDFs citing NFPA 1582 and NIOSH — bridging the gap between hazard data and preventive action.

Active pilot outreach — California departments $500 / dept pilot pricing IAFF union partnerships in progress
FastAPI · AES-256 SQLite · JWT · ChromaDB · Claude RAG · Leaflet.js

UCSC AIEA Lab

Summer Research Intern — Prof. Leilani Gilpin
Summer 2026

The Octagon

Head of Technology & Business · Built full CMS from scratch · $6,400/yr ad revenue
2023–Present

DECA President

Sacramento Country Day — competitive events, chapter operations
2025–Present

Freelance Developer

Full-stack sites, Apps Script automations, WordPress · $1,200 revenue
2024–2025
Recognition
01

USACO Platinum Division

#1 of 5,137 in Bronze. Competing at Platinum — the top global tier.

02

EMNLP Main Conference

Solo first-author submission to one of the world's most selective NLP venues.

03

3× ICML Workshop Author

Mech-Interp, Hypothesis Testing, AI for Good — all solo, 2026.

04

#1 Nationally — Commentary (JEA)

First place at the largest high school journalism competition nationally. Perfect score.

05

1st Place — Civics Unplugged Solvathon

1st of 275+ submissions. $500 cash. Selected for CU Junto cohort.

06

Codeforces Master — 2240 Rating

Competitive programming at the highest amateur level.

Contact

Let’s build
something real.

Working in AI safety or occupational health? I want to know.

siddharthshukla1010@gmail.com
LinkedIn GitHub ThePharmaBridge