B.Tech CSE @ PES University · AI / ML · Agentic Systems · Software Engineering · AI Safety Research
I build systems where machine learning, software engineering, and research meet — from autonomous ML pipelines and voice agents to internal model monitoring and mechanistic AI safety.
A systems-oriented computer science student interested in building things that are technically deep, useful, and difficult to fake.
I'm a Computer Science Engineering student at PES University, Bengaluru, working across AI/ML, agentic systems, backend engineering, and AI safety research.
My work increasingly revolves around a simple question: what happens when we stop treating AI as a black box and start building infrastructure around its internal behaviour, provenance, evaluation, and deployment?
That direction led to EdgeDaemon and its research product MechGuard, while other projects explore autonomous ML pipelines, voice agents, computer vision, forecasting, reinforcement learning, legal AI, and enterprise systems.
I enjoy working at the boundary between research and engineering — taking an idea, turning it into an experiment, then turning the useful pieces into software people can actually interact with.
A venture and research direction exploring internal evidence for AI assurance — from model development through deployment.
EdgeDaemon is the venture I am building around AI decision provenance, internal assurance, and safety infrastructure. Its current research product, MechGuard, investigates whether internal model evidence can complement conventional AI observability — from weight-space changes during fine-tuning to activation signals during deployment.
Internal assurance for AI systems — measuring internal signals, preserving provenance, and surfacing evidence for safety and security investigation.
LoRA weight-difference analysis, SVD, singular-value trajectories, subspace analysis, and checkpoint-level monitoring during fine-tuning.
Residual-stream activations, probes, cross-agent aggregation, activation alignment, and robustness evaluation for multi-agent systems.
The longer-term hypothesis connects training-time internal evidence with downstream deployment risk through model lineage and longitudinal evidence.
A broader view than the resume — production-oriented systems, research prototypes, hackathon builds, experiments, and engineering work.
Research-driven internal assurance system investigating model-internal evidence across fine-tuning and deployment. Built under EdgeDaemon.
Full-stack voice AI system being productionised for outbound calling, multilingual lead qualification, structured extraction, webhooks, and operational dashboards.
Multi-agent ML workflow connecting issue interpretation, dataset analysis, strategy generation, code generation, execution, testing, and reporting through typed orchestration.
Location-aware weather forecasting and risk-analysis system combining feature engineering, multi-horizon prediction, uncertainty analysis, and interactive visualisation.
End-to-end crop disease classification system using a fine-tuned ResNet18 pipeline with inference provenance, API services, frontend visualisation, and simulated IoT data flow.
Reinforcement-learning environment framing ML pipeline design as a sequential decision problem across preprocessing, model selection, tuning, and debugging.
Enterprise asset and resource management application built around authentication, relational data modelling, migrations, APIs, frontend workflows, and containerised deployment.
Forensic audit system reconstructing decision histories for regulated-industry AI using graph-vector memory and provenance-oriented agent workflows.
Multi-agent research orchestration concept combining streaming execution, semantic memory, web research, parallel analysis, and structured report generation.
On-device continual-learning and safety-constrained AI framework exploring quantisation, LoRA adaptation, safety interfaces, and lightweight API infrastructure.
Not everything I build lives behind a GitHub repository. These are deployed interfaces, applications, and prototypes that can be explored directly.
A separate layer of the portfolio for work that has crossed the boundary from codebase to something people can interact with.
Fully deployed weather forecasting and risk-analysis application combining the underlying ML pipeline with an interactive web interface.
Handcrafted frontend project deployed through GitHub Pages, focused on visual composition, responsive layout, typography, and polished presentation.
Visual frontend build deployed on GitHub Pages, exploring typography, composition, responsive design, and product-style web presentation.
Separate prototype / presentation dashboard for the MechGuard research system, providing a visual layer over the model-internal assurance concept.
My research direction is increasingly focused on understanding and monitoring AI systems below the input/output boundary.
The long-term goal is not to claim that we can "read model thoughts", but to build better evidence around what changes inside models and how those changes relate to downstream behaviour.
Investigating whether parameter and representation geometry changes systematically during fine-tuning regimes associated with problematic behaviour.
Exploring activation extraction, probes, cross-agent aggregation, alignment analysis, and robustness testing for multi-agent systems.
Studying whether training-time internal evidence can eventually provide useful information about downstream deployment risk.
The completed pilot validates the measurement pipeline and demonstrates substantial internal geometric movement while leaving the predictive relationship to emergent misalignment as an open research question.
A compact timeline of research, startup incubation, and engineering experience.
Building the venture around AI provenance, internal assurance, and the MechGuard research program spanning training-time and deployment-time model evidence.
Research work involving reinforcement learning, CUDA-based acceleration, reproducible evaluation, benchmarking pipelines, and AI coding-agent systems.
Selected for the founder-track incubation programme to develop EdgeDaemon, working across technical architecture, product thesis, full-stack prototyping, pitch development, and go-to-market research.
Technologies I use across research, product engineering, machine learning, backend systems, and agentic AI.
Programs, certifications, communities, and activities that have shaped the work outside the codebase.
ANTHROPIC
MICROSOFT AZURE
GOOGLE CLOUD
IIT PATNA
HACKERRANK
CODECHEF
MCKINSEY
ADOBE
Contributor across GSSoC, SSoC, and Hugging Face ecosystem projects.
Participant in the Gemini Student Ambassador programme.
Technical domain involvement across robotics and engineering activities.
Technical-domain participation and engineering community activities.
IT / technical-domain involvement within the campus community.
Competitive hockey experience representing Karnataka, bringing a team-performance perspective outside engineering.
Not every experiment belongs on the front page. This is the deeper repository layer — coursework, infrastructure experiments, frontend builds, research prototypes, and older systems.
A one-page overview of my education, experience, selected projects, technical stack, research, and engineering work.
I'm interested in internships, research opportunities, ambitious engineering teams, and people building technically difficult things in AI, ML, systems, or software.