AVAILABLE FOR INTERNSHIPS · BENGALURU, INDIA_

Yashas
Sadananda

B.Tech CSE @ PES University · Founder-track builder @ PESU Venture Labs · building traceable, auditable AI systems for regulated industries.

I build AI systems that can explain themselves. Currently founding EdgeDaemon — an AI decision-reconstruction and provenance platform — while shipping agentic pipelines, ML research systems, and safety-stack architectures across 10+ hackathons and research programs.

See EdgeDaemon → All Projects GitHub ↗
10+
Hackathons / Programs
52K+
Devs outranked (OpenEnv)
1
Founder-track startup
100+
LeetCode solved
about

Who I Am

I'm a Computer Science Engineering student at PES University, Bengaluru, working at the intersection of AI safety, agentic systems, and full-stack engineering. Most of my work centers on one question: can an AI system explain, after the fact, exactly why it did what it did?

That question became EdgeDaemon, an AI decision-reconstruction and provenance platform for regulated industries, which I'm building as a Technology Associate through PESU Venture Labs' competitive founder-track ABC program.

Outside EdgeDaemon, I've shipped five-agent trading safety stacks, RL-based ML pipeline agents, legal-defense multi-agent systems, and a ResNet18-based crop disease classifier with full inference provenance — across GitLab, Meta × PyTorch, IEEE, and FantomCode hackathons. I'm also a national-level Ice & Inline Hockey athlete representing Karnataka, which is where I actually learned how to perform under pressure as part of a team.

10+
Projects & Research Systems
6+
Certifications
8
Hackathon Finals / Shortlists
7.32
GPA · PES University
flagship venture

Building EdgeDaemon

The centerpiece of my current work — an AI decision-reconstruction and provenance platform tackling one of the hardest unsolved problems in enterprise AI: making autonomous decisions auditable after the fact.

PESU VENTURE LABS · ABC 2026 MAY 2026 — PRESENT
EdgeDaemon

An AI decision-reconstruction and provenance platform for regulated industries — healthcare, legal, and finance — where every model decision needs to be explainable, traceable, and defensible after the fact. Selected for PESU Venture Labs' Accelerating Bold Concepts program, a competitive founder-track incubation cohort, and pivoted the technical thesis from on-device continual AI personalization to decision-provenance after structured customer discovery surfaced systemic traceability failures across enterprise AI pipelines.

Architected the Retain Score — a proprietary compliance-grade benchmark quantifying catastrophic forgetting and model drift, positioning EdgeDaemon in a structurally underserved market quadrant
Validated through AML analyst stakeholder mapping, vendor landscape analysis, and forensic case studies (TD Bank's $3B AML enforcement, Dutch Benefits Scandal)
Produced regulatory landscape mapping across GDPR, SOC 2, and MAS TRM frameworks to anchor the beachhead market thesis
Built investor-facing pitch materials, Miro stakeholder ecosystem maps, and a full go-to-market strategy for cohort review
Led full-stack prototyping across React, FastAPI/Node.js, and LLM-powered agentic pipelines
Featured as centerpiece across AI safety fellowship applications, including Singapore AI Safety Fellowship
ReactFastAPINode.js LLM AgentsProvenance SystemsRegulatory Design
projects

Research & Builds

Every project here is built around the same throughline — making agentic AI systems observable, constrained, and auditable, not black boxes.

Fintech · Agentic AITEAM LEAD
AutoTrade AI
Six-layer agentic trading architecture with a 30s perceive-reason-decide-execute-self-correct loop and multi-broker integration (Zerodha, Upstox, Angel One). Led Team SynapseX for IEEE YESIST12 2026.
▸ 5-layer safety stack: confidence gating, hard constraints, circuit breakers, 2FA consent, immutable audit trail
PythonFastAPIXGBoostLSTMKafkaRedis
View on GitHub
Multi-Agent · Legal AIR2 FINALIST
Jurify
Multi-agent legal research and defense pipeline with citation-traced reasoning chains — every argument is traceable back to source documents. FantomCode 2026, Round 2 Finalist.
▸ On-device confidentiality isolation with selective memory purging for privileged documents
LangGraphClaude APIFastAPI
View on GitHub
MLOps · Agentic PipelinesGITLAB AI
ML Pipeline Orchestration Agent
5-agent autonomous ML pipeline (Trigger → Dataset Analyst → Strategist → Code Generator → Reporter) with zero manual intervention. Typed LangGraph state graphs with structured JSON observability.
▸ F1 = 0.7616 · ROC-AUC = 0.8454 · 131 automated tests passing
LangGraphClaude SonnetDockerGitLab CI/CD
View on GitHub
Reinforcement LearningTOP SHORTLIST
ML Pipeline Agent — OpenEnv
RL-based pipeline agent (preprocessing, model selection, hyperparameter optimization) using Q-learning with epsilon-greedy exploration. Meta × PyTorch Hackathon.
▸ Shortlisted from 52,000+ developers to the offline finale
Q-LearningPydanticDockerHuggingFace
View on GitHub
ML Research · ProvenanceFEATURED
MilletDataNet
End-to-end ML research system for millet leaf disease classification — fine-tuned ResNet18 with k-fold cross-validation, class rebalancing, and a full inference provenance service.
▸ Macro-F1: 0.9196 · FastAPI provenance layer · React dashboard · simulated MQTT IoT layer
PyTorchResNet18FastAPIReactMQTT
View on GitHub
Cognee · Memory SystemsPR MERGED
AuditMind
Forensic audit system reconstructing complete decision histories for regulated-industry AI. Uses Cognee's hybrid graph-vector memory to capture every reasoning step across multi-turn agent workflows.
▸ Tamper-evident provenance chains · submitted PR to topoteretes/cognee
CogneeLangGraphFastAPID3.js
View on GitHub
Full Stack · EnterpriseSOLO BUILD
AssetFlow
Enterprise Asset & Resource Management System built solo for the Odoo Hackathon 2026 virtual round — full auth, migrations, and containerized deployment from scratch.
▸ FastAPI + PostgreSQL 16 + SQLAlchemy 2.0 + Alembic + React/Vite + Docker Compose
FastAPIPostgreSQLReact/ViteDockerJWT
View on GitHub
Fraud Detection · Graph AIHACKATHON
SENTINEL
Graph-based agentic fraud detection system built for the iDEA 2.0 / Union Bank hackathon, modeling transaction networks to surface anomalous behavior.
PythonGraph MLAgentic AI
View on GitHub
TypeScript · Research AgentsMOST STARRED
NEXUS
Multi-agent research orchestration engine with streaming execution, adaptive semantic memory, and a parallel web search → analysis → structured report pipeline.
▸ Explicit memory lifecycle controls — what's stored, retrieved, and forgotten is policy-governed, not implicit
TypeScriptLangChainOpenAI API
View on GitHub
FastAPI · LangGraphPM ACCELERATOR
Weather Risk Assistant
FastAPI + LangGraph + React weather risk assistant deployed on Vercel, paired with a GradientBoosting per-location forecasting pipeline and SHAP explainability layer.
LangGraphGradientBoostingSHAPStreamlit
View on GitHub
research direction

Deep Neural Safety & Interpretability

Actively researching how language models internally generate responses — moving below the behavioral surface to understand the exact mechanisms: how specific attention heads, MLP layers, and activation circuits produce particular outputs, and what structures correspond to reasoning, memory, and decision-making.

Motivated by observed alignment failures — autonomous agents developing emergent covert communication between sub-agents, taking untracked actions, and exhibiting goal-directed behavior that bypasses intended constraints. Behavioral-interface safety alone isn't enough.

Goal: develop formal traceback methods for AI decisions at the weight/activation level — extracting a precise, mechanistic account of why a model produced a specific output, enabling proactive safety guarantees instead of post-hoc behavioral observation.

Building toward a graduate research thesis on low-level neural safety infrastructure — interpretability tooling, activation auditing, and formal provenance systems operating at the model-internals layer, not just the output layer.

MAY 2026 — PRESENT
Technology Associate Intern, PESU Venture Labs (ABC Program)
Founder-track incubation cohort. Building EdgeDaemon end-to-end — thesis, GTM, and product.
JUN 2026
AuditMind — WeMakeDevs × Cognee Hackathon
Built a traceable decision-trail reconstructor; PR submitted to the Cognee open-source repo.
APR – MAY 2026
ML Pipeline Agent — Meta × PyTorch OpenEnv Hackathon
Shortlisted to the offline finale from 52,000+ developers worldwide.
APR 2026
AutoTrade AI — IEEE YESIST12 & Jurify — FantomCode Round 2
Led a 6-layer trading safety architecture; built a multi-agent legal defense pipeline as a Round 2 Finalist.
MAR 2026
ML Pipeline Orchestration Agent — GitLab AI Hackathon
5-agent autonomous pipeline, F1 = 0.7616, fully observable via typed LangGraph state.
JAN 2026
NEXUS — Personal Research Project
Multi-agent research orchestration engine; became the most-starred repo on my GitHub.
EARLY 2026
Began AI/ML journey
Transitioned from full-stack/MERN development into serious AI and agentic systems work.
skills

Tech Stack

AI / Safety Research
LangGraphLangChainClaude API OpenAI APIRAGAgentic AI LoRA/QLoRADPOContinual Learning Audit Trail DesignMech. Interp. (foundational)
ML / Modeling
PyTorchXGBoostLSTM Q-LearningResNet18SHAP
Languages
PythonTypeScriptJavaScript C++CJavaSQL
Full-Stack & Infra
ReactNext.jsFastAPI Node.jsTailwindDocker GitLab CI/CDHuggingFace Spaces
Data Layer
PostgreSQLMongoDBMySQL SupabaseFirebaseRedisKafka
CS Foundations
DSA (100+ LC)OOPOS Computer NetworksComputer ArchitectureDBMS
credentials

Certifications & Activities

Claude Fluency in AI & Claude 101
ANTHROPIC
Forward Learning Programme
MCKINSEY
Generative AI & Neural Networks
MICROSOFT AZURE
AI/ML Program
IIT PATNA
Python (Advanced)
HACKERRANK
C Programming
CODECHEF
Open Source Contributor
GSSoC · SSoC · HUGGINGFACE
National Ice & Inline Hockey Athlete
TEAM KARNATAKA

Let's Build

Looking for AI engineering / AI safety / full-stack internships at teams building serious things. If you're working on something that needs to be provably correct — I'd love to talk.

yxshas565@gmail.com