C0
PRANAV_PAPPU
Pranav Pappu
Builder / AI Researcher
/scroll ↓
Activity
Contributions · 2026github ↗
1,265 contributions in 2026
JanFebMarAprMayJunJulAugSepOct
Mon
Wed
Fri
LessMore
About
C1ABOUT
Year 2 CS student at NUS. I build AI-powered tools and ship fast.
Currently:
  • Building AssessMate for early childhood intervention assessors as a CCSGP Fellow.
  • Building Tweakler.
  • Research Assistant under Prof. Ooi Wei Tsang on dynamic 3D Gaussian Splatting benchmarking.
Projects
C3ASSESSMATE
motorlanguagesocialcognitiveself-careNOTE_03sessions →
  • AI case-notes copilot for structured documentation.
  • Piloted with two assessors to measure documentation-time reduction.
  • Built for PDPA-sensitive child-development data.
/site ↗
C5LOCAL SMALL LM
attn · causal×12next tokHello,I'malangmodel
  • Built a custom GPT-2 124M variant from scratch in PyTorch on local hardware.
  • Tokenizer, transformer blocks, optimizer, mixed-precision training, and FineWeb-Edu evaluation.
  • Trained on 10B tokens.
/github ↗
C7EVOSERVE
↑ qualitydiversity →
  • Harness-quality and evaluation control plane for coding-agent systems.
  • Compares policy candidates and independently verifies captured-run evidence.
  • Produces auditable quality and regression reports.
/wip
Experience
C9WINS
1STReplit Agent Hackathon (Singapore)
  • Built Remember — an AR smart glasses simulator for dementia patients with real-time face recognition, contextual information overlay, and a personal knowledge graph powered by live speech extraction.
/github ↗
2NDGoogle DeepMind × Cactus Compute Hackathon
  • Built SecureClaw — a privacy-first on-device AI assistant using FunctionGemma with smart cloud fallback, PII detection, and local-vs-cloud routing.
/github ↗
TOP 30%NVIDIA Nemotron Model Reasoning Challenge
  • Used a MAP-Elites-style search over fine-tuning recipes, then trained a rank-32 attention LoRA on 14,400 solver-generated reasoning traces for Nemotron-3-Nano-30B.
  • Placed in the 70th percentile in a field of 5,223 participants.
—Hack&Roll 2026
  • Built JustPromptBro — a VSCode extension with event-driven keystroke tracking, real-time behavioural feedback, and a Solana smart contract for on-chain game logic.
/github ↗
—Jane Street Kaggle Competition
  • Built a neural network from scratch and leveraged XGBoost for time-series forecasting on unlabelled market data to predict prices 6 months ahead.
Contact