Projects

Side projects, research prototypes, and passion builds — from LLM simulation engines to neuromorphic hardware models.

🧪

Pulp AI

LLMsFastAPIpgvectorRedisDocker

A simulation platform that lets product teams test copy, features, and pricing against AI-generated customer personas — shipping decisions in minutes instead of weeks. Built on FastAPI, Postgres (pgvector), Redis, and Docker.

pulp.ai/simulate
👤

Sarah Chen, 34

Product Manager at a Series B SaaS startup

Key concerns

  • Price sensitivity at renewal
  • Feature overwhelm during onboarding
  • Unclear ROI for enterprise tier

AI recommendation

Lead with a 3-step onboarding wizard and surface ROI metrics in the first week.

Sample output from the LLM persona engine

🐛

Pest Detection System

Computer VisionEfficientNetResNetData Augmentation

Trained EfficientNet, ResNet, and Inception models for pest identification on cotton and rice farms. Advanced data augmentation improved pipeline robustness by 14%, directly helping 10,000+ farmers take corrective action without compromising yield.

Model Accuracy Comparison

EfficientNet-B4
94.2%
ResNet-50
88.7%
Inception-v3
86.1%

+14% improvement from advanced data augmentation pipeline

📈

CPI Forecasting across Indian Markets

LSTMARIMATime SeriesEconomics
GitHub →

Scraped ~1.2M CPI price points from 300+ Indian markets and trained LSTM models that reduced 30-day-ahead MAE by 18% versus ARIMA, revealing inflation hot-spots across eight economic zones.

LSTM beats ARIMA by ~87% MAE

Academic & Research Projects

🧠

Thalamo-Cortico-Thalamic Relay on SpiNNaker

Neuromorphic ComputingSpiNNakerComputational Neuroscience

Modelled the TCT relay system on the SpiNNaker neuromorphic platform for low-cost, low-power cortical signalling. Built under Prof. Basabdatta Sen Bhattacharya at BITS Pilani.

🎮

Multi-Agent Reinforcement Learning for Cooperative Hunting

MARLGame TheoryNash EquilibriumReinforcement Learning

Built a predator–prey game with Cooperate/Kill actions and trained agents with Nash-Q learning. Demonstrated that learned policies converge to the game's Nash equilibrium, confirming stable cooperative dynamics.

View on GitHub →
🌐

Neural Machine Translation — Verb Dependencies

NLPSeq2SeqMachine Translation

Built a customised Seq2Seq variant to improve NMT translation accuracy by reducing the proximity in verb phrases between source and target languages. Research conducted at IIIT Hyderabad.

View on GitHub →