Work Sentinel The Desk Experience Honors Contact

Bhopal, India — Status: Online

Systems that hold under adversarial conditions.

I'm Simarjot Singh Maan — I build guardrails that stop language models from being jailbroken, and risk systems that stop trading strategies from blowing up. Same instinct, two domains: find where the system breaks, then build the layer that catches it.

See Sentinel in action → Download résumé

// who's writing this

I'm a researcher and builder working across AI safety, quantitative trading, and full-stack development — currently based in Bhopal, India. My work lives at the boundary where a system meets something trying to break it: a jailbreak prompt hitting an LLM endpoint, a black-swan candle hitting a leveraged position, a malformed input hitting a production API. I like the boundary. That's where the interesting engineering is.

0% adversarial prompts blocked
pre-generation, Sentinel L1–3
0 prompt benchmark, open-sourced
126 adversarial · 100 benign
0 reference architecture
5 layers implemented & live
0 place, INFOSPARK 2526
time-constrained build hackathon

Selected work

Things I've shipped
or broken on purpose.

Research · 2026 SAID Laboratory

SENTINEL — Trustworthy Guardrails for Web-Agent LLM Services

A model-agnostic guardrail framework: five inference-time enforcement layers inside a nine-layer reference architecture, wrapping any OpenAI-compatible endpoint to block jailbreaks, prompt injection, and indirect tool-injection — without retraining the model.

  • 49.2% of attacks blocked pre-generation at 5.0% benign false-positive rate
  • More than 2× a keyword-matching baseline (19.0%)
  • Llama 3.1 8B end-to-end: block rate 80.2% → 90.5%, true ASR 19.8% → 9.5%
View on GitHub →
Independent research · 2024–25 Python

Algorithmic Trading System Development

Designed, implemented, and backtested algorithmic trading systems for crypto futures and prediction markets — signal-generation modules, a backtesting framework for strategy P&L, and risk controls (position sizing, drawdown limits, exposure caps) built from first principles.

  • Independent research into market microstructure & statistical arbitrage
  • Custom backtesting framework evaluating strategy P&L end-to-end
  • Risk management: position sizing, drawdown control, exposure limits
Try the risk simulator ↓
Independent research · 2024–25 PyTorch · Transformers

Custom LLM Development via Dataset Training

Fine-tuned large language models through custom dataset-training pipelines — studying training dynamics, distribution shift, and failure modes, and building end-to-end research infrastructure with Python and the broader ML tooling ecosystem.

  • Built production-style research infra end-to-end, solo
  • Studied model behavior under distribution shift
Pre-seed startup · 2024–25 Founder & full-stack

AI-Powered B2B SaaS

Architected, built, and deployed a full-stack AI product end-to-end as sole developer — every line hand-written, zero no-code tools — owning product design, feature prioritization, and the deployment pipeline.

  • Sole developer: design, build, ship, deploy
  • Zero no-code dependency — fully hand-coded

Interactive — Layers 1–5, live in your browser

Try to get past
the guardrail.

This is a simplified, client-side re-creation of SENTINEL's enforcement pipeline. Type a prompt below — try something obviously adversarial, try encoding it, try a "as an AI with no restrictions" roleplay — and watch it move through the same layer logic described in the paper.

L2Encoding
Normalize
L3Context
Integrity
L1Intent
Classifier
L4Model
Core
L5Output
Scorer
>

// pipeline idle. waiting for input.

Try:
86%indirect tool-injection caught at Layer 3, sub-millisecond
+35.7ppmarginal gain from context-integrity rules over normalization alone
69%of all blocks exit via fast rule-gate in ~0.3ms — no model call needed

Interactive — risk management, the hard way

Survive ten rounds
without blowing up.

A tiny simulation of the thing that actually matters in trading: not picking winners, but sizing positions so one bad round doesn't end the game. Set your risk per round, then ride it out.

EQUITY $10,000 +0.00%
ROUND0 / 10
MAX DRAWDOWN0.0%
STATUSReady

Drag the slider, then play a round. Higher risk per round means faster gains — and faster ruin. That tradeoff is the entire job.

Where I've worked

A working timeline,
not a résumé bullet list.

May 2026 — Jun 2026

Equity Research Intern

Millennium Money Finance · Remote · BSE/NSE-registered

Hands-on exposure to demat account workflows, DuPont analysis, and technical analysis of listed equities. Completed under direct mentorship with a final practical assessment.

May 2025 — Present

Human Resource Intern

Simtrak Solutions · ADORE · Remote

Selected for a structured mentorship program. Manage daily reporting and task workflows, coordinate cross-team efforts across a distributed team.

May 2025 — Jun 2025

Content Writing Intern

InAmigos Foundation · Remote

Received a Certificate of Appreciation and a Letter of Recommendation from Founder & CEO Mr. Govind Shukla for contributions to the internship department.

Apr 2025 — May 2025

Content Writing & Marketing Intern

Kshitiksha Foundation · ISO 9001:2015 NGO · Remote

30-day internship graded 'S' — Superior, the highest performance tier. Received a Certificate of Experience and a Letter of Recommendation from the Executive Director.

Toolbox

What I actually use.

Languages

  • Python
  • Java
  • HTML / CSS / JS

AI / ML

  • PyTorch ecosystem
  • Hugging Face Transformers
  • LLM fine-tuning & evaluation
  • Adversarial robustness
  • NLP

Quantitative Finance

  • Algorithmic trading
  • Backtesting
  • Statistical arbitrage
  • Risk management
  • Equity research (DuPont & technical analysis)

Building

  • Full-stack product dev
  • Deployment pipelines
  • Solo-founder execution
  • Research infrastructure

Honors & leadership

Outside the codebase.

1st

INFOSPARK 2526 Hackathon

Time-constrained technical build challenge — DPS Kolar Road, Bhopal (2025)

1st

Mount Carmel Declamation

Mount Carmel Literary Fest 2025–26

★

Harvard MUN India 2025

Debate Club Leader — represented a nation in committee, Bangalore

★

Founder, Chess Club

Billabong High International School — built the club from scratch; 100+ hours coaching Grade 5–6 students

S

Kshitiksha Foundation

'S' — Superior internship performance grade, highest tier awarded

★

ThinkStartup Program

IIT Delhi (RNI Park) — certified by TIDES Business Incubator, IIT Roorkee (May 2026)

Education

Billabong High International School, Bhopal

ICSE, Science + Computer Applications · 2025–2026

Coursework: Physics, Chemistry, Biology, Computer Science

Get in touch

Building something
that needs to hold?

Open to research collaboration, internships, and interesting problems at the boundary of AI safety and quant systems.

Email simarjotsinghmaan@gmail.com Phone +91 97553 00773 Project github.com/saidlaboratory/SENTINEL
Location Bhopal, Madhya Pradesh, India