Hello, I’m
AI researcher& systems engineer.
Hyderabad, India · 17.3850° N, 78.4867° E
CurrentlyLLM reliability at LLM Lens · cyber security at AICTE
I study how language models fail, then build systems that can detect, explain, and act on uncertainty—from hallucination probes to cyber-intelligence platforms and autonomous machines.
A short edit of the work that best represents how I think: research depth, system constraints, and something real at the other end.
Research · LLM Lens
I study whether hallucination signals can be read from transformer hidden states—and what a system should do when those signals say an answer is unreliable.
Two arXiv preprints spanning internal representation probes and black-box behavioral detection.
Autonomous surveillance rover
A hybrid edge–cloud rover that combines LiDAR, depth sensing, GPS, computer vision, autonomous navigation, and human takeover for perimeter monitoring.
Led the system from hardware integration to field testing, a first-prize finish at Prakalp 2025, a conference presentation, and a published Indian patent application.
Cyber threat-intelligence platform
An analyst-facing system for detecting AI-generated misinformation, mapping coordinated narratives, explaining risk, and preserving an auditable decision trail.
Built with the winning Smart India Hackathon 2025 team and now continuing beyond the hackathon under the AICTE Cyber Security Cell.
Multimodal route optimization across road, rail, sea, and air with emissions forecasting, uncertainty ranges, explainable trade-offs, and shipment-level audit trails.
A privacy-first desktop workspace with persistent memory, grounded document Q&A, streaming chat, local models, and a Qdrant-backed retrieval layer.
AIS and simulated SAR fusion for anomaly detection and dark-vessel correlation, with the limits of synthetic data documented as part of the design.
My current work asks two complementary questions: what can internal activations tell us, and what can model behavior reveal when those internals are closed?
“
Detection is only the first step. The useful system is the one that knows whether to answer, verify, retrieve evidence, or abstain.
Current research directionDo hallucination signals transfer across datasets?
What survives a change in model family or generation?
Can a lightweight controller intervene before unreliable text reaches a user?
A weak-supervision framework that labels generations using grounding signals, then trains five probe architectures directly on LLaMA-2-7B hidden states so detection no longer needs external verification at inference time.
Open on arXivA black-box detector built from semantic entropy, embedding geometry, chain-of-thought consistency, and paraphrase stability—useful when model internals are unavailable.
Open on arXivMy best work happens at the seam between a research question and a usable system: datasets, models, evaluation, APIs, interfaces, and deployment all informing one another.
Research engineering, applied AI, backend, and ambitious product teams.
LLM Lens
Building hidden-state extraction, probe training, transfer evaluation, and calibration pipelines for LLM hallucination detection across datasets and model families.
AICTE Cyber Security Cell
Continuing TattvaDrishti beyond SIH: literature grounding, misinformation analysis, source verification, and productization for real analyst workflows.
Recurse, KMIT
Designing technical problem statements, evaluating builds, and mentoring students through system design, Git, and implementation issues.
OpenAI ChatGPT Lab · India
Selected among 28 students nationwide for the inaugural cohort exploring how students use LLMs for learning, research, and collaboration.
I’m pursuing computer science at KMIT and data science at IIT Madras at the same time. The dual degree is useful context; the work is the point.
I gravitate toward systems where an intelligent model has to survive contact with constraints—limited compute, noisy sensors, uncertain data, real users, or a deadline that refuses to move.
Away from the desk, the same curiosity shows up in aerospace, cars, astronomy, books, and an entirely reasonable amount of Star Wars.
Education
BS · Data Science and Applications
B.Tech · Computer Science and Engineering
Selected by function, not logo count.
PyTorch · Hugging Face · Transformers · LLM evaluation · OpenCV · scikit-learn
Next.js · React · Vue · FastAPI · Flask · Node.js · WebSockets
PostgreSQL · Redis · Celery · Docker · Linux · Qdrant · CUDA
Jetson Nano · Pixhawk · ArduPilot · MAVLink · LiDAR · ESP32
Have a difficult system in mind?
I’m open to research collaborations, internships, deep-tech teams, and product work where AI has to be more than a demo.