Computer Science Graduate Student · Dublin, Ireland

Software engineer building ML systems, from reinforcement-learning trading agents to production APIs.

MSc Computing (AI & NLP) student at Dublin City University, seeking Summer 2027 software engineering internships in ML, backend, and systems.

3.2

Sharpe Ratio (RL agent)

50K+

Ticks/sec handled

92%

Circuit ID accuracy

2×

Hackathon 1st places

Engineering with product clarity.

I build reliable systems and interfaces, from ML pipelines to production web apps, with an emphasis on measurable outcomes and maintainable architecture.

I'm Sri Krishna Adithya, a computer science graduate student with hands-on experience across reinforcement learning, full-stack development, and real-time systems. My work spans training RL agents in simulated market environments, building low-latency data pipelines, and shipping production-facing web apps end-to-end.

I enjoy owning problems from architecture through deployment: defining the approach, building it, and measuring whether it actually works. I'm actively seeking Summer 2027 software engineering internships in ML, backend, and systems.

Postgraduate

MSc Computing in Artificial Intelligence & NLP

Dublin City University · Incoming, Sept 2026

Focus: AI, NLP, and applied machine learning

Undergraduate

Bachelor of Computer Applications

SRM Institute of Science and Technology, Vadapalani · Graduated May 2026

Focus: Data structures & algorithms, computer networks, operating systems, database systems. GPA 8.1

Engineering case studies

Selected work with problem context, technical architecture, and measurable results.

2025

Reinforcement Learning Market-Making Agent

GitHub

RL-based market-making agent trained in a custom Level-2 order-book simulation, benchmarked against a classical quoting baseline.

Problem

Classical market-making strategies like Avellaneda-Stoikov rely on fixed stochastic assumptions that don't adapt to changing order-book conditions, leaving PnL and inventory risk on the table.

Architecture

Custom OpenAI Gym environment simulating Level-2 order-book dynamics, dynamic quoting, inventory controls, and transaction-cost modeling. PPO and DQN agents trained against the environment with reward shaping tied to inventory risk and realized PnL.

Technologies

  • Python
  • OpenAI Gym
  • Stable-Baselines3
  • PPO
  • DQN
  • NumPy
  • Pandas

Engineering decisions

  • Modeled transaction costs and inventory penalties directly in the reward function to keep the agent from overtrading.
  • Compared PPO and DQN under identical environment seeds to isolate algorithm choice from environment variance.
  • Benchmarked against the Avellaneda-Stoikov closed-form strategy as a classical baseline.

Outcomes

  • Achieved a Sharpe ratio of 3.2 and ~35% higher PnL than the Avellaneda-Stoikov baseline in simulation.
  • Held inventory within ±10% of target even under high-volatility scenarios, indicating strong generalization.

2025

Real-Time Trading Dashboard

GitHub

Low-latency dashboard streaming and visualizing live market data with custom analytics and role-based auth.

Problem

Retail-facing trading tools often lag on tick-level updates and lack composable analytics, making it hard to react to fast-moving markets.

Architecture

WebSocket ingestion pipeline backed by optimized Python/C++ components, Redis caching for hot reads, and a React frontend with D3/Recharts visualizations.

Technologies

  • Python
  • C++
  • WebSockets
  • Redis
  • React
  • D3.js
  • Recharts

Engineering decisions

  • Split latency-critical tick processing into C++ components while keeping orchestration in Python for iteration speed.
  • Cached derived analytics (VWAP, rolling volatility) in Redis to avoid recomputing them on every tick.
  • Added role-based authentication to separate read-only viewers from accounts with alerting configuration access.

Outcomes

  • Handled 50,000+ ticks per second through the WebSocket pipeline in load testing.
  • Supported 1,000+ concurrent users during load tests without degrading dashboard responsiveness.

2024

TOR Network Forensics & Deanonymization Dashboard

GitHub

Network-forensics platform that studies TOR traffic patterns to identify potential entry-exit node pairs.

Problem

Understanding TOR traffic-correlation risk requires tooling most security researchers don't have readily available, and existing approaches are fragmented across relay metadata and traffic analysis.

Architecture

Relay metadata collected and categorized via the Onionoo API. A correlation engine compares packet timing and size patterns to flag candidate entry-exit node pairs, paired with an interactive topology visualization and a reporting module.

Technologies

  • Python
  • Onionoo API
  • Network Analysis
  • React
  • Data Visualization

Engineering decisions

  • Used packet timing and size correlation rather than payload inspection, staying within TOR's traffic-analysis threat model rather than attempting decryption.
  • Built the topology visualization interactively so relay relationships could be explored rather than read off a static report.
  • Generated reporting summaries separately from the live correlation engine to keep analysis reproducible.

Outcomes

  • Achieved 92% accuracy identifying circuits in simulation.
  • Generated correlation-summary reports with a mean turnaround of 2.5s.

2025

AgroScope: Agricultural Supply Chain & Matchmaking Platform

GitHub

Platform connecting farmers with agri-tech startups through data-driven, location-aware matching.

Problem

Farmers and agri-tech startups lack a data-driven way to find each other, relying on informal networks that don't account for crop type, geography, or startup fit.

Architecture

React/TypeScript frontend, Node.js backend, and a FastAPI microservice running a Random Forest model for crop categorization and match ranking. MongoDB for storage, with Haversine geo-scoring for location-aware recommendations.

Technologies

  • React
  • TypeScript
  • Node.js
  • FastAPI
  • Python
  • scikit-learn
  • MongoDB

Engineering decisions

  • Split the ML-driven matching logic into a separate FastAPI microservice so the Node.js backend could stay focused on CRUD and auth.
  • Used Haversine geo-scoring to weight matches by real-world proximity instead of naive region matching.
  • Trained a Random Forest classifier for crop categorization, prioritizing interpretability for non-technical users over marginal accuracy gains.

Outcomes

  • Delivered real-time, location-aware match recommendations connecting farmers with relevant agri-tech startups.
  • Built a type-safe frontend with dashboard views for both sides of the marketplace.

Planned / in progress

Planned

Planned / In Progress

Real-Time Fraud Detection Pipeline

Streaming ML pipeline concept to score transactions in real time with explainable risk signals for analysts.

Problem

Legacy batch scoring introduces multi-minute detection lag, increasing chargeback exposure on high-volume payment flows.

Architecture

Event-driven pipeline: Kafka ingest → feature store (Redis) → ONNX inference service → alerting API, with a Next.js/WebSocket dashboard for live incident review.

Technologies

  • Python
  • FastAPI
  • Kafka
  • Redis
  • ONNX
  • PostgreSQL
  • Next.js
  • Docker

Engineering decisions

  • Planning to use ONNX over raw PyTorch serving to minimize inference latency on CPU-only nodes.
  • Partitioning features by entity ID to keep hot-path reads O(1) and avoid cross-tenant leakage.
  • Designing a shadow-mode deployment path so new models can be validated without blocking production traffic.

Outcomes

  • Target: real-time scoring well under the multi-minute lag of batch systems.
  • Target: meaningfully reduce false positives while preserving high recall on held-out fraud labels.
  • Target: surface ranked explanations in the ops UI to speed up analyst triage.

Planned

Planned / In Progress

Domain-Specific RAG Research Assistant

Retrieval-augmented system concept for querying technical papers with citation-grounded answers and an evaluation harness.

Problem

General-purpose chat models hallucinate citations and struggle with domain-specific notation in CS research PDFs.

Architecture

Ingestion workers chunk and embed documents into a vector store. Query path: hybrid retrieval (BM25 + vector) → reranker → constrained generation with source spans returned to the client.

Technologies

  • TypeScript
  • Next.js
  • Python
  • LangChain
  • Vector DB
  • OpenAI API
  • PostgreSQL

Engineering decisions

  • Planning parent-child chunking so answers can cite precise paragraphs while retrieval uses broader context windows.
  • Building an offline eval set with citation-accuracy scoring to run before each prompt or index change.
  • Designing embedding writes to be idempotent so re-indexing is resumable after ingestion failures.

Outcomes

  • Target: materially improve citation accuracy over a baseline single-vector RAG setup.
  • Target: keep end-to-end query latency low enough for interactive use, including reranking.

Planned

Planned / In Progress

Distributed Systems Visualizer

Interactive teaching-tool concept to simulate consensus, replication lag, and partition behavior in distributed databases.

Problem

Students struggle to reason about failure modes in consensus protocols when learning from static slides alone.

Architecture

Simulation engine in Rust (WASM) to drive state transitions; React UI to render node graphs and timelines. Scenarios as declarative JSON configs loaded at runtime.

Technologies

  • Rust
  • WebAssembly
  • React
  • TypeScript
  • Canvas API
  • Vite

Engineering decisions

  • Planning to compile the core simulation to WASM to keep step updates smooth with many virtual nodes.
  • Separating scenario definitions from engine logic so new labs can be authored without recompiling.
  • Persisting learner progress locally with exportable session logs for grading.

Outcomes

  • Target: usable in a real distributed-systems course to make failure-mode behavior tangible.
  • Target: open-source the engine so scenario authoring isn't locked to one course.

Technical stack

Tools and domains I use regularly in coursework, research, and production-oriented projects. Hover a skill for context.

Languages

Python
Primary language, used in every real-world project
C++
Latency-critical components in the trading dashboard
JavaScript
Comfortable across Node.js and browser environments
TypeScript
Default for frontend and backend work since 2024
Python
Primary language, used in every real-world project
C++
Latency-critical components in the trading dashboard
JavaScript
Comfortable across Node.js and browser environments
TypeScript
Default for frontend and backend work since 2024

AI / ML

Reinforcement Learning
Core focus, trained PPO/DQN agents to a 3.2 Sharpe ratio
Stable-Baselines3
Used to train and benchmark PPO/DQN agents
scikit-learn
Trained the Random Forest model behind AgroScope's matching
OpenAI Gym
Built a custom Level-2 order-book environment in Gym
Stochastic Modeling
Applied to order-book and transaction-cost modeling
LOB Simulation
Simulated realistic Level-2 order-book dynamics
Reinforcement Learning
Core focus, trained PPO/DQN agents to a 3.2 Sharpe ratio
Stable-Baselines3
Used to train and benchmark PPO/DQN agents
scikit-learn
Trained the Random Forest model behind AgroScope's matching
OpenAI Gym
Built a custom Level-2 order-book environment in Gym
Stochastic Modeling
Applied to order-book and transaction-cost modeling
LOB Simulation
Simulated realistic Level-2 order-book dynamics

Frontend

React
Primary frontend framework across every shipped project
Vite
Used for fast local dev in the AgroScope frontend
Tailwind CSS
Used to build this site and the AgroScope UI
Radix UI
Used for accessible, unstyled UI primitives
Recharts
Used for live market-data visualization
SVG-based Visualization
Built the interactive TOR relay topology view
React
Primary frontend framework across every shipped project
Vite
Used for fast local dev in the AgroScope frontend
Tailwind CSS
Used to build this site and the AgroScope UI
Radix UI
Used for accessible, unstyled UI primitives
Recharts
Used for live market-data visualization
SVG-based Visualization
Built the interactive TOR relay topology view

Backend & Data

Node.js
Backend runtime for AgroScope and the trading dashboard
Express.js
REST API layer behind AgroScope's backend
FastAPI
Served the ML matching microservice behind AgroScope
REST APIs
Built REST endpoints during my Astra Dyne Global internship
WebSockets
Streamed 50K+ ticks/sec in the trading dashboard
MongoDB
Cut MongoDB response times ~20% at Astra Dyne Global
PostgreSQL
Comfortable with relational schema design and SQL
Redis
Cached derived analytics to avoid recomputing on every tick
Drizzle ORM
Type-safe query layer for Postgres-backed projects
Node.js
Backend runtime for AgroScope and the trading dashboard
Express.js
REST API layer behind AgroScope's backend
FastAPI
Served the ML matching microservice behind AgroScope
REST APIs
Built REST endpoints during my Astra Dyne Global internship
WebSockets
Streamed 50K+ ticks/sec in the trading dashboard
MongoDB
Cut MongoDB response times ~20% at Astra Dyne Global
PostgreSQL
Comfortable with relational schema design and SQL
Redis
Cached derived analytics to avoid recomputing on every tick
Drizzle ORM
Type-safe query layer for Postgres-backed projects

Dev Tools

Git
Daily version control across every project
GitHub
Where all my real, shipped work lives
JWT Auth
Implemented role-based access in the trading dashboard
Bcrypt
Used for password hashing alongside JWT auth
Docker
Used to containerize local dev environments
Git
Daily version control across every project
GitHub
Where all my real, shipped work lives
JWT Auth
Implemented role-based access in the trading dashboard
Bcrypt
Used for password hashing alongside JWT auth
Docker
Used to containerize local dev environments

Experience

Internship experience and competitive-hackathon results.

  1. Software Engineering Intern

    Jul 2025 to Oct 2025

    Astra Dyne Global

    • Optimized MongoDB queries and data models, helping reduce API response times by around 20%.
    • Built and refined REST endpoints and contributed to backend and frontend integration for new features.
    • Worked within production constraints on a live codebase, covering clean-code practices, debugging, and real-world data edge cases.

Hackathon achievements

  • 1st Place at TechnoVera'24 (SRM Vadapalani)
  • 1st Place at Origin (SIMATS)
  • 2nd Place at Invente'24 (SNUC/SSN)
  • Special Mention at Hackathrone (VIT)