AI & DATANIT AGARTALA2029

Aman
Banik.

Engineering high-performance compute, predictive ML, and agentic architecture.

Aman Banik
fig_01 / aman_banik
ML/DL • High-Performance Compute • GenAI

6+

+projects shipped

FlyRank AI

ML Internship ↗
01 /

Projects and workflows

01

FlyRank AI Capstone

PREDICTIVE MODELING

Engineered a transparent heuristic baseline and compared it against Machine Learning models (Naive vs. Tuned Random Forest) to predict organic traffic decay. Employed GroupKFold cross-validation, aggressive regularization, and relative feature engineering to combat overfitting on power-law distributed web data, ultimately prioritizing 5,562 pages for editorial triage with a robust 71% Precision@100.

Feature ExtractionModel Eval & ComparisonHyperparameter TuningReportLabScikit-LearnRandom Forest
Data
Model
Eval
Report
02

CUDA_0 (Gargantua & Gray-Scott)

HIGH-PERFORMANCE COMPUTE

Rewrote CPU-based relativistic black hole rendering into parallel Numba CUDA kernels, scaling particle count to 42,400+ and boosting FPS from 9 to 153 on an RTX 5060 (16x speedup). Implemented a parallel numeric PDE solver sustaining 3.7 billion state updates/second.

PythonNumbaCUDA
CPU
9 FPS
RTX 5060
153 FPS
42,400+ Particles / PDE Cells
03

Neko Agent

DISTRIBUTED AI SYSTEMS

Smart Retail B2B Procurement Guardrail designed as a 4-agent distributed system (EDA, Analyst, Strategist, Guardrail) via FastAPI microservices. Enforces hard purchase budgets and integrates with live PostgreSQL databases and Google Drive MCP.

Google ADKFastAPIMCP
Guardrail
EDA
Analyst
Strategist
04

The Grounded1

HIPAA-COMPLIANT AGENTS

Two-agent system preventing LLM hallucinations in healthcare access. A Root Agent plans actions while a Filter Agent enforces HIPAA standard operating procedures before execution, achieving 1-2s latency and 100% compliance.

PythonGemini APIJSON Policy
Root Agent
Plan Action
HIPAA Filter
Verify Policy
Execute
05

Moogle Translate

NEURAL MACHINE TRANSLATION

Seq2Seq English-to-Hindi translator trained on 130K pairs. Vectorized sequence feeding to directly utilize Tensor Cores, cutting epoch time by 90%. Added Luong Global Attention to scale vocab to 20K words.

TensorFlowKerascuDNN
Training Loss
Vanilla Seq2Seq
Fast Vectorized
Global Attention
06

COVID-19 Time Series Forecasting

PREDICTIVE MODELING

Built and compared 3 forecasting models (Prophet, ARIMA, SARIMA) using walk-forward validation. Prophet achieved highly stable results (MAPE 2-6%).

ProphetARIMAStatsmodels
Prophet Model
Predicted Actual
02 /

Capabilities

MACHINE LEARNING

  • TensorFlow / Keras
  • Scikit-Learn
  • LangChain / Agentic AI
  • Deep Learning / GenAI

DATA & COMPUTE

  • Python / C
  • Numba / CUDA HPC / cuDNN
  • Pandas / NumPy
  • SQL (T-SQL)

SYSTEMS & BACKEND

  • FastAPI
  • Google Gemini API / MCP
  • Git & Version Control
  • Linux / WSL2
03 /

About

Engineering student at NIT Agartala — building high-performance computational systems and production-grade agentic AI, not just academic experiments.

My work sits at the intersection of parallel computing, machine learning, and hardware optimization. I care about systems that exploit extreme throughput (CUDA) and multi-agent reasoning to solve complex constraints.

Looking Forward with Vision

aman@portfolio ~

$ python profile.py

Loading Aman Banik...

{
  "role": "AI/HPC Researcher & ECE Engineer",
  "university": "NIT Agartala",
  "status": "actively_plotting_and_crafting",
  "domain_stack": [
    "Predictive Machine Learning & Deep Learning",
    "High-Performance Computing",
    "Semiconductor Physics, Analog Systems, Discrete Math & Linear Algebra"
  ],
  "hobbies": [
    "making silicon sweat with custom compute kernels",
    "engineering agentic systems that actually execute",
    "optimizing data science workflows for bare-metal hardware",
    "fixing systems without completely destroying them"
  ],
  "mission": "Building autonomous AI ecosystems, optimized from the circuit level up."
}
                        

$ _

04 /

Contact

Open to conversations —
reach out and let's talk compute.

X (TWITTER) @Alph_ed