Aditya / AI × Systems × Quant

Building / Researching

I buildintelligentsystems.

Computer Science & Data Analytics student at IIT Patna building AI/ML systems, quantitative research pipelines, and production-grade data applications.

Explore my work GitHub Resume
Aditya seated in a woven chair on a high-rise balcony, working on a laptop, with a waterfront skyline behind him.
Fig. 00Subject — Aditya
Education
B.S. CSDA · IIT Patna ’29
Now
Building INDRA · SIH 2026
Based
Patna, India
  1. Data
  2. Signal
  3. Model
  4. Decision
[01]The Signal

I like turningmessy datainto systems.

Most of what I build sits where models meet decisions: a regime classifier that decides how capital is allocated, a forecaster that decides how much stock to reorder, an agent that decides which Python to run against a dataset. The model is one step. The system around it — data, evaluation, deployment, feedback — is the work.

Black-and-white side profile of Aditya against a dark wall, eyes closed, holding a glass.
Fig. 01Signal / noise
  • 01AI / MLAgents, fine-tuned LLMs, gradient boosting, explainability.DataScout · IndiEye
  • 02Quantitative ResearchRegimes, alphas, backtests that respect time.Regimes · WorldQuant BRAIN
  • 03Data SystemsStreaming pipelines from raw reports to verified decisions.INDRA · DemandIQ
  • 04Cloud InfrastructureServerless AWS — Lambda, S3, API Gateway, DynamoDB.DataScout
  • 05MLOps & InferenceShipping, serving and running models efficiently — learning now.Now
  • 06Developer ToolsAPIs, dashboards and interfaces that make models usable.FastAPI · Streamlit
[02]Selected Work

Selectedwork

Eight systems across data platforms, quant research, ML platforms, agentic AI and interpretability. Each one is a pipeline, not a notebook — click through for the case study.

Now building01 / 08Data Platform · SIH 20262026 – Present

INDRA

Intelligent National Disaster & Weather Platform

Turns fragmented citizen reports, official warnings and weather observations into verified, explainable weather events for India’s emergency operations centres. I build the backend.

  1. Reports + feeds
  2. Redpanda
  3. Clean · geocode · dedup
  4. DBSCAN + H3
  5. Evidence
  6. Receipt
  7. Verified event
  • Python
  • FastAPI
  • PostGIS
  • Redpanda (Kafka)
  • Redis
  • H3
  • DBSCAN
  • SeaweedFS (S3)
  • Docker
Commits · #1 contributor
691
Tests passing
1,788
Hazard types
16
Factor verification receipt
7
Case study Private team repository
02 / 08Quant Research2025

Market Regime Detection Framework

Hybrid HMM + K-Means regimes on NIFTY 50

Hidden Markov Models and K-Means label the market’s regime; an XGBoost classifier predicts it; the strategy follows the regime.

  1. NIFTY 50
  2. Features
  3. HMM + K-Means
  4. XGBoost
  5. Allocation
  • Python
  • HMM
  • K-Means
  • XGBoost
  • MLflow
  • Streamlit
Market regimes
3
NIFTY 50 history
15y
Engineered features
28
Regime prediction accuracy
78%
Case study Source private
03 / 08ML Platform2026

DemandIQ

Retail Demand Forecasting & Replenishment Engine

Forecasts next-week sales for every SKU in every store — then turns the forecast into a reorder quantity and a risk alert.

  1. Data
  2. Feature Engineering
  3. Forecast
  4. Inventory Risk
  5. Alert
  • Prophet
  • XGBoost
  • FastAPI
  • Streamlit
  • MLflow
Pipeline layers
6
Day lag features
7·14·28
Service level target
95%
Forecast models
2
Case study Source
Flagship04 / 08AI System2026 – Present

IndiEye

Market Intelligence Platform

Institutional-grade market sentiment for retail investors: news, social chatter and filings, read by a fine-tuned LLM and lined up against price action.

  1. News + Social + Filings
  2. NLP
  3. Sentiment
  4. Price Action
  5. Insight
  • Next.js
  • FastAPI
  • PostgreSQL
  • Llama 3
  • AWS
Source streams
3
Fine-tuned for sentiment
Llama 3
Open-source services
2
Case study ML service News service
05 / 08AI System2026

DataScout

Agentic AI Data Analyst

Ask a dataset a question in plain English. An agent writes the Python, runs it in a sandbox and shows its work.

  1. User
  2. Natural Language
  3. AI Agent
  4. Python Execution
  5. Analysis
  6. Insight
  • Amazon Bedrock
  • AWS Lambda
  • S3
  • API Gateway
  • DynamoDB
  • Streamlit
Team — led
4
Encryption at rest
AES-256
SQL required
0
Case study Source
06 / 08AI System2026

Sanjivani AI

Multimodal crisis intelligence for Bihar floods

Reads distress posts, satellite imagery and supply history in one pipeline — triage from text, flood extent from pixels, resource forecasts per district.

  1. Posts + satellite + history
  2. NLP triage
  3. Flood segmentation
  4. Resource forecast
  5. District dashboard
  • DistilBERT
  • U-Net
  • YOLOv8
  • XGBoost
  • FastAPI
  • Streamlit
  • PostGIS
  • Docker
Modalities: text · imagery · tabular
3
API endpoints
5
Tests passing
34/34
Case study Source
07 / 08Agent Systems2026

Agentic AI Core

A foundation layer for autonomous agents

An LLM planner decomposes the goal, an executor runs tools with validation, timeouts and retries, an evaluator checks the result, and memory carries context forward.

  1. Goal
  2. Planner
  3. Executor
  4. Evaluator
  5. Result
  6. Memory
  • Python
  • Pydantic v2
  • FAISS
  • Chroma
  • structlog
  • FastAPI
  • Swappable LLMs
Agent roles
4
Built-in tools specified
7
Design specs incl. threat model
8
Case study Private · in design
08 / 08Research2026

LLM Pruning & Explainability

Do explanations survive compression?

Pruning makes models cheaper; explanations make them trustworthy. This study measures what pruning does to the explanations.

  1. Model
  2. Pruning
  3. Sparsity
  4. Explanation
  5. Faithfulness
  • SHAP
  • Integrated Gradients
  • DistilBERT
  • RoBERTa
Sparsity range
40–80%
Pruning methods
3
Peak Hessian vs 0.73 baseline
14,090
Case study

Low-magnitude weights go first. Explanations stay faithful — reported up to 80% sparsity.

More builds04
[03]System Map

How I thinkabout systems

Every project on this page is the same loop with different data. Hover, tap or tab through the nodes.

Node 01 · Data

Start where the mess is.

  • Citizen reports, official warnings, METAR observations
  • Financial market data — NIFTY 50, VIX
  • Retail demand — Walmart M5 sales, prices, calendar
  • News, social chatter and filings

Seen in

  • INDRA
  • Regime Detection
  • DemandIQ
  • IndiEye
  • DataScout
[04]Experience

Where I’vebuilt

  1. 2026Sep 2026 – Present

    Smart India Hackathon 2026

    Team Sixth Sense

    Core Platform Engineer (Backend) · INDRA

    • Build and maintain INDRA’s core platform (layers 1–3 and 5–8a): ingestion, processing, geo-analytics, event fusion, data platform and the real-time API, for problem statement SIH26069 — National Weather Big Data Analytics.
    • Streaming intake with a transactional outbox into Redpanda, pollers for official warnings and weather observations, a dead-letter queue and an S3-compatible data lake.
    • A deterministic seven-factor Verification Receipt over PostGIS, H3 and DBSCAN clusters, with a hash-chained audit ledger and role-based access on every write.
    • #1 contributor — 691 commits; 1,788 tests passing in the latest recorded run.
    • FastAPI
    • Redpanda
    • PostGIS
    • H3
    • Event fusion
  2. 2026Feb 2026 – Mar 2026

    AI for Bharat

    Patna, Bihar

    AI Engineer (Developer)

    • Built DataScout, an agentic AI data analyst for natural-language data analysis on Amazon Bedrock and AWS serverless (S3, Lambda, API Gateway, DynamoDB).
    • Implemented LLM-generated Python execution for analytical results — more accurate and auditable than generated text; built the Streamlit interface.
    • Designed the security model: IAM role isolation, AES-256 encryption, audit logging and sandboxed execution.
    • Led a team of 4 into the Prototype Development Phase of a national AI hackathon.
    • Amazon Bedrock
    • Agents
    • Serverless
    • Team lead
  3. 2026Jan 2026 – Feb 2026

    WorldQuant BRAIN

    Remote

    Alpha Research Trainee

    • Designed and simulated equity alpha signals in WorldQuant BRAIN’s Fast Expression language on the US TOP3000 universe, with market and industry neutralization.
    • Applied decay, truncation and delay constraints to manage turnover and risk; evaluated with Sharpe ratio, turnover, drawdown and out-of-sample test periods.
    • Alpha Research
    • US TOP3000
    • Fast Expression
    • Neutralization
    • Backtesting
  4. 2025Mar 2025 – Jun 2029

    Indian Institute of Technology Patna

    Patna, Bihar

    B.S. Computer Science & Data Analytics

    Certification · NISM Series VIII — Equity Derivatives Certification Examination

[05]Now

Rightnow.

Building one system in production, and going deeper on how models get shipped, served and run efficiently.

Building

INDRA backend

The core platform for a national weather-event verification system, for Smart India Hackathon 2026.

  • Streaming ingestion with an outbox, DLQ and data lake
  • Geo-clustering with PostGIS, H3 and DBSCAN
  • An explainable verification receipt and audit ledger
Case study

Learning

Inference engineering

How models actually run — memory, latency and cost — and how to make them cheaper and faster.

  • Running local LLMs with Ollama and llama.cpp
  • Quantization (GGUF, 4/8-bit) and its accuracy trade-offs
  • Serving with vLLM: continuous batching and the KV cache
  • Measuring tokens/s, time-to-first-token and memory

Learning + building

MLOps

Making models reproducible, versioned and shippable — not just trained.

  • Experiment tracking and a model registry with MLflow
  • Containerised model services with Docker and FastAPI
  • CI for ML code with GitHub Actions and pytest
  • Data and model versioning with DVC; monitoring with Prometheus and Grafana

Practising

Fine-tuning

Adapting open models to narrow domains without retraining them from scratch.

  • Llama 3 fine-tuned for financial sentiment (IndiEye)
  • Parameter-efficient fine-tuning: LoRA and QLoRA with PEFT
  • Building instruction datasets and evaluating adapters
[06]Research / Thinking

Beyondshipping.

Shipping a model is half the job. The other half is knowing when it’s right, why it works, and where it breaks.

N-01Quantitative finance · IEEE-format paper

Market Regime Detection

QuestionCan an unsupervised view of market state make a supervised allocator more robust?

Methods
Hidden Markov Models + K-Means (hybrid labelling); XGBoost with sample weighting
Data
15 years of NIFTY 50 · 28 engineered features (realized vol, VIX dynamics, skew)
Regimes
Low-volatility · Trending · Crisis
Result

78% regime prediction accuracy. In backtest, regime-driven allocation delivered 14.2% CAGR vs 10.8% buy-and-hold, with max drawdown cut from 50% to 18%.

N-02Interpretability · Model compression

LLM Pruning & Explainability

QuestionWhen a transformer is pruned, do its explanations stay faithful?

Methods
Random, L1 unstructured, L1 structured pruning · 40–80% sparsity
Explainers
SHAP · Integrated Gradients · FCor faithfulness metric
Models / Data
DistilBERT, RoBERTa · IMDb, Yelp
Result

Magnitude pruning preserves faithfulness up to 80% sparsity. Random pruning produces high-curvature landscapes (Hessian up to 14,090 vs 0.73 baseline) that break SHAP’s linearity assumptions.

N-03Systematic equities · WorldQuant BRAIN

Alpha Research

QuestionWhat survives neutralization, decay and delay — and what was just noise?

Universe
US TOP3000 equities
Tooling
Fast Expression language · market & industry neutralization
Constraints
Decay · truncation · delay
Result

Signals evaluated on Sharpe ratio, turnover and drawdown, with out-of-sample test periods as the final check.

[07]Stack

Thetoolchain

9 groups, 58 tools — the ones that show up in the work above, plus the two I’m going deep on now (marked).

01

Languages

05
  • Python
  • C++
  • SQL
  • JavaScript
  • Rust
02

ML & Quant

08
  • XGBoost
  • Prophet
  • HMM
  • K-Means
  • SHAP
  • Llama 3
  • MLflow
  • Backtesting
03

Data

05
  • PostgreSQL
  • DynamoDB
  • Pandas
  • NumPy
  • Scikit-learn
04

Web

05
  • FastAPI
  • Streamlit
  • Plotly
  • Next.js
  • React
05

Cloud

07
  • AWS
  • S3
  • Lambda
  • API Gateway
  • DynamoDB
  • Bedrock
  • IAM
06

Tooling

06
  • Git
  • Docker
  • pytest
  • Alembic
  • yfinance
  • Uvicorn
07

Streaming & Geo

06
  • Redpanda (Kafka)
  • PostGIS
  • H3
  • DBSCAN
  • Redis
  • SQLAlchemy (async)
08

MLOps

08

Learning + building

  • MLflow tracking & registry
  • Docker Compose
  • GitHub Actions
  • DVC
  • ONNX
  • Model serving (FastAPI)
  • Prometheus
  • Grafana
09

Inference

08

Learning

  • Ollama
  • llama.cpp
  • GGUF quantization
  • vLLM
  • KV cache & batching
  • LoRA / QLoRA (PEFT)
  • Hugging Face Transformers
  • ONNX Runtime
Portrait of Aditya in a charcoal three-piece suit and paisley tie against a clear sky.
Fig. 08Identity
[08]About

I build at the intersection ofmachine learning, data,markets and production systems.

I’m a Computer Science & Data Analytics undergraduate at IIT Patna. My work sits where models meet real decisions — detecting market regimes, forecasting retail demand, extracting sentiment from news and filings, and building agents that compute answers instead of guessing them.

I care about the unglamorous parts: features that don’t leak the future, evaluation that matches how a model will actually be used, and interfaces that let someone act on the output. Right now that means building the backend of INDRA, a disaster-intelligence platform for Smart India Hackathon 2026, while going deeper into inference engineering and MLOps.

Education

B.S. Computer Science & Data Analytics

Indian Institute of Technology Patna

Mar 2025 – Jun 2029

Interests
  • AI / ML
  • Quantitative Research
  • Data Infrastructure
  • MLOps
  • Inference Engineering
  • Cloud Systems
  • Applied Research
[09]Open Source / Code

Read thesource

The public half of the work lives on GitHub. Some projects — INDRA (a team repository), the regime framework and IndiEye’s core — are private for now.

~/aditbytes $ ls --sort=featured○ snapshot · live stats unavailable
Explore GitHub github.com/aditbytes
[10]Contact

Build somethingworth shipping.

I’m most interested in problems where data, models and markets meet — research collaborations, messy datasets, or a system that needs to get from a notebook to production. Email is the fastest way to reach me.

Send an email GitHub LinkedIn Resume

System ready

The nextsystemcould be yours.Let’s talk