MACHINE LEARNING,MARKETS & THE MIND.

AI & Data Science · CMRIT ’28 · Bengaluru

Rahul J. builds learning systems for retrieval, ranking, anomaly detection and vision, and studies the economics and psychology that decide whether such systems matter.

Currently: Founding AI & Product Engineer, Gojocart. Reading for a B.E. in Artificial Intelligence & Data Science, CMRIT, class of 2028.

Scroll

I work where disciplines meet. Engineering supplies the method, statistics the discipline, and economics and psychology the questions worth asking; each project is an attempt to make one of them answer to the others.

// Profile

Selected work

01 / 05

NorthStar

An offline-first, retrieval-augmented tutor for low-connectivity classrooms, answering on a CPU in under two seconds.

  • RAG
  • FAISS
  • Local LLM

1st place · CodeRush 2026

OpenENV

A multi-objective decision agent that triages, schedules and executes enterprise work under real-time constraints.

  • Agents
  • FastAPI
  • Docker

0.98+ Easy/Medium · 0.90+ Hard

TalentMatch AI

Explainable candidate ranking on SBERT embeddings and fifteen engineered signals, with per-candidate reasoning.

  • NLP
  • Ranking
  • SBERT

+22% ranking correlation vs TF-IDF

SentinelSense

Real-time anomaly detection on multivariate sensor streams, combining classical detectors with an LSTM autoencoder.

  • Time series
  • LSTM
  • IoT

F1 0.87 on SKAB · −31% false positives

ValoVision

A computer-vision pipeline that detects events in FPS gameplay and turns them into measures of player behaviour.

  • YOLOv8
  • OpenCV
  • Vision

91% event-detection precision

Drag, scroll or use the arrow keys

One inquiry, several instruments.

// Practice

My work moves between machine learning, quantitative markets and the study of human behaviour. The subjects differ; the habit of mind does not. Each asks how a system should decide under uncertainty, and how to tell a real signal from a convenient story.

I learn by building. Theory is tested against a working artefact, results are measured against an honest baseline, and every model is expected to explain itself.

The long-term aim is independence of thought and work that compounds: a self-directed curriculum across deep technology, economics and behaviour, applied to problems that scale.

A trajectory in four movements.

// Trajectory

  1. Strategy and pattern

    Foundations

    Early study of strategic play through chess and programming.

    Learned Pattern recognition and foresight are general instruments; they transfer across every serious domain.

  2. Quantitative trading

    Markets

    Traded forex through proprietary-firm evaluations using rule-based, risk-first algorithmic strategies.

    Learned Markets punish emotional bias. A modest system run with discipline outperforms a brilliant one run on instinct.

  3. Formal study, CMRIT

    2024 – 2028

    B.E. in Artificial Intelligence & Data Science: machine learning, probability, linear algebra, algorithms and database systems.

    Learned Mathematics turns raw aptitude into engineered systems.

  4. Industry, Gojocart

    2025 – present

    Founding AI & Product Engineer at a hyperlocal quick-commerce start-up based in Cassino, Italy, working remotely.

    Learned A model is only as useful as the operation it changes.

Method and result.

// Projects, in depth

  1. 01 · NORTHSTAR 1st place · CodeRush 2026

    NorthStar: an offline AI tutor

    Python · SentenceTransformers · TF-IDF · FAISS · local LLM

    Question

    Can a capable study assistant work where the internet does not?

    Method

    Offline retrieval-augmented generation: FAISS and SBERT semantic search feeding a 4-bit quantised LLaMA-family model for on-device inference.

    Result

    Answers in under two seconds on a CPU. First place, Education domain, CodeRush 2026, a 24-hour national hackathon.

  2. 02 · OPENENV Meta AI OpenEnv Hackathon

    OpenENV: an enterprise decision agent

    Python · FastAPI · Pydantic · Docker · OpenEnv API

    Question

    How should an agent balance email triage, scheduling, task execution and system health when they compete?

    Method

    An RL-inspired policy layer that finds the weakest performance dimension at each step and chooses constraint-aware actions, keeping decision logic separate from LLM execution.

    Result

    Baseline scores of 0.98+ on Easy and Medium tasks and 0.90+ on Hard. Deployed with Docker to Hugging Face Spaces behind FastAPI and Gradio, following OpenEnv’s reset, step and state contract.

  3. 03 · TALENTMATCH

    TalentMatch AI: explainable ranking

    Python · scikit-learn · SentenceTransformers · NLP

    Question

    Can candidate ranking be both more accurate and easier to justify?

    Method

    Cosine similarity on SBERT embeddings combined with fifteen-plus engineered features, from TF-IDF vectors to skill-match ratios, in a weighted ensemble tuned by grid search.

    Result

    A 22% improvement in ranking correlation over a TF-IDF baseline, measured with NDCG@10 and Precision@K, with LLM-assisted, per-candidate explanations.

  4. 04 · SENTINELSENSE

    SentinelSense: anomalies in sensor streams

    Python · scikit-learn · LSTM

    Question

    How do you catch genuine faults in multivariate IoT data without drowning operators in false alarms?

    Method

    Isolation Forest, One-Class SVM and a sliding-window LSTM autoencoder, with dynamic thresholds from rolling z-scores and adaptive percentiles, and severity-scored alerting.

    Result

    False positives down 31% against fixed thresholds. On the SKAB benchmark, F1 of 0.83 for Isolation Forest and 0.87 for the LSTM autoencoder on held-out data.

  5. 05 · VALOVISION

    ValoVision: reading gameplay as behaviour

    Python · OpenCV · YOLOv8

    Question

    Can raw FPS footage be turned into reliable measures of how a player behaves?

    Method

    Frame differencing, contour analysis and a YOLOv8 model fine-tuned on more than 1,200 labelled frames, stabilised by sliding-window majority voting.

    Result

    91% event-detection precision on a held-out set, feeding kill-zone frequency, positional heatmaps and reaction-latency analysis.

Instruments and inquiries.

// Disciplines

Engineering

  1. Python, SQL and C
  2. PyTorch, scikit-learn, NumPy and pandas
  3. NLP and vision: Hugging Face Transformers, SentenceTransformers, NLTK, OpenCV
  4. Retrieval and serving: FAISS, FastAPI, Docker
  5. Cloud: AWS EC2, S3 and Amazon EKS on Linux

Inquiry

  1. Probability, statistics and linear algebra
  2. Algorithms and data structures
  3. The economics of artificial intelligence
  4. Psychology, sociology and decision-making
  5. Strategy, markets and real-world systems

Present inquiries.

// Now

  1. Building

    • Forecasting and recommendation for hyperlocal merchants
    • A body of applied ML work in retrieval, agents and vision
    • Deeper fluency in algorithms and data structures

  2. Studying

    • Automation economics and its effect on labour
    • Behavioural models of judgement under uncertainty
    • Graduate management study, beginning with the GMAT

  3. Founding

    • A student initiative applying behavioural psychology to resilience and strategic thinking
    • Curriculum and support structure in design

Practice and peer review.

// Experience and service

  1. 2025 – PRESENT

    Gojocart

    Founding AI & Product Engineer · Cassino, Italy (remote)

    Forecasting

    Recommendation and demand-forecasting prototypes for hyperlocal merchant inventory, built on collaborative filtering and historical sales patterns.

    Operations

    Automated order routing, merchant onboarding and analytics dashboards, cutting manual processing by an estimated 40%.

    Intelligence

    Market intelligence from web data aggregation and NLP-based competitor signals, informing pricing and geographic expansion.

  2. 2026

    AICCoNS, International AI Conference

    Research paper reviewer

    Scope

    Submissions in machine learning, NLP, computer vision, federated learning, AI security and adversarial AI, assessed for correctness, novelty, methodology and reproducibility.

  3. 2026

    EAMCON, via Microsoft CMT

    Conference reviewer

    Scope

    Submissions in federated learning, intelligent computing and smart applications, assessed for originality, technical depth and practical impact.

  4. HONOURS

    Recognition and credentials

    Competitions · certifications

    First place

    CodeRush 2026, a 24-hour national hackathon (Education domain); Think Build Showcase, CMRIT project exhibition.

    Competition

    Meta AI OpenEnv Hackathon: multi-objective, RL-inspired task-automation agent.

    Certification

    Architecting Solutions on AWS (AWS, edX); Python Project for Data Engineering (IBM, edX); The Economics of AI (University of Virginia, Coursera).

// Contact

Correspondence.

Open to research collaboration, internships and problems that sit between fields.

Bengaluru, India