Shreyash Poddar
I work on AI security and interpretability, and build the software around them, from Android apps to quantum backends.
B.Tech student, SRM Institute of Science and Technology, Chennai.
- cgpa
- 9.92
- github
- ShreyashPoddar
- shreyash-poddar
scroll
nvidia a100 · sxm4 · 80 gb hbm2e
The kind of hardware my research trains on. Scroll to take it apart.
01experience
a91f3c2 (HEAD -> main)
Technical Director, ECHO SRM
08/2026 – present
Leading the team that officially builds websites for SRM students, including SEPS 2.0, the capstone-project portal used by thousands of students at once.
4d27b10
Software Development Intern, NeevCloud
04/2026 – 07/2026
Got the internship by placing at the Fastathon hackathon. Built AuraHR, an AI hiring platform: resume matching, AI video interviews with camera proctoring, coding tests and automatic candidate emails.
7c0e9aa
Manager, Cherry+ Network Club
09/2024 – 05/2026
Ran the technical side of the club's events.
02skills
What I work with
- languages
- Python, C++, C, Java, TypeScript, JavaScript, SQL
- ai / ml
- PyTorch, scikit-learn, NLP, Transformers, RAG, jailbreak robustness, XAI
- nlp & llms
- tokenization, embeddings, attention, sampling, KV cache, post-training (SFT, RLHF, DPO)
- interpretability
- probing, sparse autoencoders, activation steering, activation patching, logit lens
The grid is attention for one sentence: each tile's height is how much one word looks at another. The raised row is "instruction".
- security
- threat modeling, OWASP LLM Top 10, MITRE ATT&CK, CVE / CVSS, prompt injection, red teaming
- software
- React, Next.js, Node.js, FastAPI, PostgreSQL, Redis, Android, Flutter
- cloud
- AWS, Docker, Linux, CI/CD, CUDA
- quantum
- Qiskit, PennyLane, IBM Quantum, quantum machine learning
03projects
10/2026
LLM Firewall
A small model that checks text before an AI agent reads it and flags hidden instructions. It was tested only on attack sources it never saw in training. On instructions hidden inside emails it scored F1 0.88, against 0.57 for a widely used open-source detector.
passed 0blocked 0
09/2026
PAA, Personal Assistant Agent
An Android assistant that turns WhatsApp and Telegram messages into reminders. It understands English and Hinglish, works hands-free, and runs its AI model on the phone itself, so chats never leave the device.
08/2026
SafeSlope-NER
A landslide warning system: field sensors, a risk score for every slope, a live map, and incident reports over WhatsApp. Every photo is checked by AI and then confirmed by a person before an alert goes out.
04research
10/2026 [confidential]
JEPA-based local AI model
A language model small enough to run on a laptop, built as a clean test bed for AI-safety experiments such as jailbreaks and unlearning. I built the evaluation side. Training and evaluation run on NVIDIA DGX GPUs.
09/2026 [confidential]
Explainable AI for fair loan approval
Does explaining a loan model's decisions change how accurate and how fair it is? A pre-registered study on three credit datasets, including 2.18 million mortgage applications.
The bars are an illustration of feature attributions, not results from the study.
08/2026 [confidential]
ANRF-funded ML research
Government-funded machine learning research under the Anusandhan National Research Foundation. The details stay redacted until publication.
04/2026 [confidential]
Quantum-native backends
Running a backend's security checks as quantum circuits, tested in simulation and on real IBM quantum hardware, over 450,000 shots.
Above: a dilution refrigerator, the gold "chandelier" that cools a quantum chip to near absolute zero.
bell pair · (|00⟩ + |11⟩) / √2
Before measurement neither qubit has a definite state on its own. Measure them and they always agree.
Each qubit alone has no definite state.
05recognition
- 04/2026Fastathon hackathon, placed; earned the NeevCloud internship
- 2025 – 2026Merit Scholarship, SRM, for a perfect 10.0 CGPA in second year
- 2024 – 2026Highest SGPA in section, ECE: a perfect 10.0 in 3 of 4 semesters
- 09/2025Tech Mesh Grand Prix (Swap Edition), 3rd place
- 08/2024Best Presentation, Pythoneers, ACM SIGCHI SRM
- volunteerDare to Help NGO, Varanasi: 30 hours of on-ground social work