Observe the system.
Find structure in sequential data, financial signals, and real-world telemetry.
Applied AI researcher / builder
Mithilesh Adhinarayanan. State.Move.Result.
I’m Mithilesh Adhinarayanan.
I turn noisy observations into better decisions — then build the systems that make the decision useful.
CSE · AI & DS / SASTRA Hosur, India
01 observe → model → act ∞
State-space observatory001 / ∞
One cube. A universe of decisions.4.32 × 1019legal 3×3 states
Reinforcement learning✳Quantitative finance✳Applied machine learning✳Data + decisions✳Reinforcement learning✳Quantitative finance✳Applied machine learning✳Data + decisions✳
01The through-linecuriosity → computation
Different systems.
The same question.
How do you turn uncertainty into a better next move?
Financial time series, traffic queues, network threats, a scrambled cube. I’m drawn to problems where mathematics explains the structure, learning finds the signal, and engineering makes the answer usable.
At SASTRA, I work across AI research and end-to-end applications: from a first experiment to data pipelines, model backends, and the interface someone actually uses.
Choose with intent.
Connect reinforcement learning, probability, and constraints to decisions.
Build. Measure. Iterate.
Take models beyond notebooks into APIs, applications, and working systems.
02Selected systemsideas, made executable
Beyond the notebook.
Systems I’m building at the intersection of models, data, and people.
01 / Adaptive networkArchitecture sketch
ObserveLearnConstrainAct
Adaptive Traffic
Management
From fixed timers to a network that responds. Camera telemetry feeds DQN agents, emergency preemption, and dynamic routing — with a hard safety layer learned policies cannot override.
43 testsdocumented in the repository
Edge → APIMQTT, Redis, FastAPI + dashboard
Problem, approach & current scope
Problem: traffic demand changes faster than fixed signal schedules. Approach: per-intersection policies, a coordinator, telemetry contracts, and mandatory clearance constraints. Scope: an end-to-end simulated platform with documented camera and controller integration points; real-road validation is a separate step.
02 / Decision simulatorOffline-first / AI-ready
AERIS
An offline-first drone and counter-drone training simulator with degraded sensors, explainable AI cues, transparent scoring, and after-action review.
4-step looptrack / classify / alert / resolve
Local-firstsynthetic scenarios, no external API
Problem, approach & current scope
Problem: training decisions should remain measurable when sensors are noisy or degraded. Approach: deterministic scenario packs, a browser radar console, a local classifier cue, and an end-of-run review. Scope: a synthetic training prototype for the SIH 2026 brief.
03 / Human-first AIPipeline sketch
CraftHaat
A voice note and a photo become a bilingual artisan catalog. Offline-first capture meets self-hosted speech, image, and language models.
Offline-firstcapture now, synchronize later
Voice → catalogWhisper + Ollama + ONDC preparation
Problem, approach & current scope
Problem: catalog entry should not require typing or a reliable connection. Approach: a local draft queue, self-hosted inference, and structured catalog generation. Scope: a prototype; production authentication and live ONDC gateway publishing are ongoing work.
04 / Decision environmentRemediation sketch
Exchange Rate OpenEnv
Before a model can make a good decision, its inputs need care. An agent chooses whether to accept, replace, or drop missing, spiking, or stale currency-feed observations.
3 actionsaccept / patch / drop
3 task levelsmissing data / spikes / latency
Problem, approach & current scope
Problem: corrupted ticks contaminate downstream quantitative models. Approach: explicit observation/action spaces and increasingly difficult remediation tasks. Scope: a simulated feed-remediation environment, distinct from the exchange-rate forecasting research below.
05 / Semantic earth dataEvent fusion sketch
EFSAT
Event-based Fusion for Semantic Satellite Analytics and Transmission. Search observations in natural language, review ranked change events, compare evidence, and prepare compact semantic packets instead of repeatedly moving full-resolution imagery.
Event-firstranked changes with provenance
Offline-readysynthetic data, no external API
Problem, approach & current scope
Problem: analysts should be able to find meaningful changes without repeatedly transferring entire image collections. Approach: natural-language search, a ranked change-event queue, an AOI evidence panel, and semantic packet preview. Scope: an offline-first prototype using synthetic demonstration data.
06 / Entity resolutionCandidate graph
Business Entity Resolution
An end-to-end matching pipeline that turns messy business records into ranked candidate pairs: country-aware blocking, TF-IDF and token features, a gradient-boosted matcher, and a threshold tuned for the challenge objective.
Two-stagecandidate generation → matching
F0.5 tunedthreshold selected on validation
Problem, approach & current scope
Problem: the same entity can appear under different names, addresses, and source systems. Approach: reduce the search space with country and token-aware blocking, then score candidates with name, address, and country features. Scope: a challenge pipeline that writes the required candidate and matching outputs.
07 / Bayesian market statesPosterior sketch
PyMC Stochastic Volatility
A Bayesian volatility proof of concept for EUR/USD returns. A latent Gaussian random walk tracks changing variance while Student-T observations accommodate heavy-tailed market shocks.
Latent staterandom-walk volatility process
Interpretableposterior checks with ArviZ
Problem, approach & current scope
Problem: market returns do not have constant variance, especially during shocks. Approach: infer a latent volatility process with a Gaussian random walk and Student-T observation model, then inspect convergence and posterior behavior. Scope: a preparatory Bayesian time-series notebook for a PyMC-focused GSoC path.
Also on the workbench
experiments / foundations / open sourceDeep Learning Lab
Foundational models and notebooks. PyTorch, TensorFlow, and NPTEL coursework.
PyMC / GSoC preparation
Bayesian modelling and open-source preparation. A learning and contribution track.
AI + web applications
Data pipelines, model workflows, and application backends built end to end.
Original repository index
Browse the broader set of projects, experiments, and preparation tracks on GitHub.
More original systems
research tools / applied ML / open source08Physics + AI
ML4Sci Shower Classifier
A GSoC 2026 preparation project for Machine Learning for Science. The notebooks build deep-learning pipelines for calorimeter-shower classification with PyTorch CNNs, ROC-AUC evaluation, and SHAP heatmaps that make the model’s physical evidence inspectable.
Open repository ↗09Routing PoC
Dynamic Closure Routing
A FastAPI microservice proof of concept that ingests temporary road closures and changes graph edge weights so Dijkstra routing recalculates detours immediately, without rebuilding a full planetary index.
Open repository ↗10Practice system
Prep / DSA + CP
A personal preparation repository for data structures, algorithms, Codeforces, competitive programming, puzzles, and the structured practice that supports quantitative and systems work.
Open repository ↗03Research notebookquestions still in motion
Learn from the unexpected.
Two active directions at SASTRA: uncertainty in markets and long-running threats in networks.
DNN–RL exchange-rate forecasting
Combining deep neural networks and PPO policy-gradient reinforcement learning with crude oil, gold futures, and NIFTY 50 as macroeconomic signals.
Mar 2026 → present
Results & publication status
The research reports predictive accuracy above 95%. The paper is submitted to IEEE and awaiting peer review. This reported result needs to be read with the study’s metric, data split, and baseline; a public evaluation artifact is not linked here yet.
Advanced persistent threat detection
Investigating deep learning and anomaly detection to identify stages of sophisticated, long-term threat activity hidden in enterprise network traffic.
Feb 2026 → present
The question behind the work
How can models connect subtle anomalies across network activity rather than treating every event in isolation? This is an ongoing research direction, with investigation focused on threat stages and persistent behavior.
04Capability atlasfrom theory to deployment
Tools for every move.
A complete view of my working toolkit, organized around what it helps me build.
01 / Learning
AI, ML & deep learning
Learn a representation. Optimize a policy.
Reinforcement learning · deep learning · PPO · DQN · multi-agent RL · neural networks · policy gradients · backpropagation · transfer learning · time series · model optimization
Applied in Adaptive Traffic ↗02 / Language
NLP & language models
Turn unstructured language into useful systems.
NLP · LLMs · Transformers · BERT / GPT architectures · text classification · sentiment analysis · tokenization · embeddings · prompt engineering · retrieval-augmented generation
Applied in CraftHaat ↗03 / Perception
Computer vision
Pixels become observations, observations become action.
CNNs · object detection · image classification · image segmentation · feature extraction · real-time video processing · vehicle detection
Camera-to-policy telemetry ↗04 / Languages
Programming foundations
Express the idea. Understand the machinery.
Python (advanced) · Java / Swing / GUI · C++ · C · SQL · JavaScript · data structures and algorithms · competitive problem-solving
Explore code across the stack ↗05 / Applications
Web & full-stack engineering
Connect the model to the person using it.
Frontend and backend development · HTML / CSS · REST and WebSocket APIs · database integration · offline-first mobile workflows
Explore application work ↗06 / Data
Data science & analytics
Clean inputs. Better experiments.
Feature engineering · data cleaning · exploratory data analysis · statistical analysis · data visualization · financial and sequential data pipelines
The research notebook ↗07 / Markets
Quantitative finance
Find a signal. Account for uncertainty.
WorldQuant BRAIN · custom expression languages · alpha research · algorithmic trading · backtesting · exchange-rate modelling · financial forecasting · macroeconomic analysis · risk analysis · portfolio optimization
WorldQuant research experience ↗08 / Infrastructure
Tools, systems & security
Make experiments reproducible and systems connected.
High performance computing · Linux · Git / GitHub · Docker · Raspberry Pi · Redis · MQTT · Jupyter Notebook · Google Colab · VS Code · network security · APT detection · anomaly detection
From edge telemetry to APIs ↗09 / Foundations
Mathematics & problem solving
Look for the invariant beneath the complexity.
Advanced mathematics · probability theory · linear algebra · calculus · statistical modelling · optimization · Olympiad mathematics · physics · proofs and problem-solving
Mathematics in public ↗05Journey + milestoneslearn / make / compound
Small moves.
Compounding progress.
Apr 2026 → present
WorldQuant / BRAIN
Research Consultant
Developing and backtesting quantitative alphas using probability and statistical modelling. Reached Gold Genius level while researching signal quality across global markets.
Conditional Consultant track / Apr–May 2026
Jun → Jul 2026
XYlofy AI / Bengaluru
AI & Data Science Intern
Built and deployed machine-learning architectures for automation projects, connecting Python data pipelines, REST APIs, and model backends in an Agile development workflow.
Feb 2026 → present
SASTRA Deemed University
Student Researcher
Investigating APT detection since February and DNN–RL financial forecasting since March. Connecting anomaly detection, sequential learning, and decision-making.
Research directions ↗2026 → present
Freelance / remote
AI/ML Developer
Delivering custom machine-learning solutions and full-stack applications: data pipelines, training workflows, and deployment-ready backend APIs.
Freelancer profile ↗06Outside the notebooka little more of me
The instinct to solve.
Speedcubing is where state spaces become something you can hold. Patterns, constraints, algorithms — and the satisfaction of a better next move.
01 / The classic
3×3×3 cube
A small object with an enormous state space. Each legal turn changes the permutation; the challenge is choosing a useful sequence.
8! × 3⁷ × 12! × 2¹⁰
43,252,003,274,489,856,000 reachable states.
A visual playground for 2×2, 3×3, 4×4, 5×5, Pyraminx, and Clock. Official competition results are highlighted below.
3×3 cube37.70s
2×2 cube13.17s
Pyraminx6.61s
07Education + community
Always a student.
SASTRA
Deemed University
B.Tech / Computer Science Engineering
Artificial Intelligence & Data Science
Thanjavur, Tamil Nadu
Yale / Coursera
Financial Markets certification
NPTEL SWAYAM
Deep Learning coursework
Harvard CS50W
Web Programming with Python & JavaScript
Mathematics, in public.
Contributing to Mathematics, Computer Science, and Physics communities on Stack Exchange. Sharing solutions and engaging with difficult technical questions.
Stack Exchange ↗AoPS & problem solving.
Engaging with challenging mathematics and CS competition problems through Art of Problem Solving since 2022. A place to develop intuition, rigor, and better explanations.
Art of Problem Solving / 2022 → ongoingLearning in publicSmall commits.
Small commits.
Long-term curiosity.
Experiments, research preparation, and working systems — the workbench keeps moving.
Current activity on GitHub ↗
Open GitHub profile ↗
08The next movelet’s build something useful
Open to meaningful problemsGood questions.
gmail.com
Good questions.
Working prototypes.
Research collaborations, internships, and freelance AI/ML work. Especially where learning systems meet financial data or real-world decisions.
mithilcuber@+91 9488694227Hosur, India