Building

precise

systems.

Full NFT marketplace on a custom Solidity smart contract. The dual-confirmation escrow is the most technically interesting part — funds lock in the contract on purchase. Both buyer and seller must independently confirm before transfer. If either cancels, buyer is refunded and asset returns to seller. Enforced by code. No intermediary, no trust required.

• ERC-721 standard with custom VaultItem struct tracking complete on-chain lifecycle state
• Dual-confirmation escrow: buyer + seller must both confirm before fund release
• AES-GCM client-side encryption before IPFS upload — content mathematically private until purchased
• Real-time event system via blockchain events — no polling; Zustand auto-sync loop (3s interval)
• Ethers.js v6 with custom FetchRequest headers to bypass ngrok browser warning in dev
• Preview URI system for marketplace browsing without decryption — full mint / list / cancel / buy / burn

Decentralized protocol for verifiable humanitarian impact. The core innovation: Proof of Benevolence (PoB) — a novel consensus primitive that validates real-world charitable activity on-chain. Unlike PoW (energy) or PoS (capital), PoB treats verified impact as a first-class blockchain primitive. Enables ESG funds to operate with mathematical proof rather than self-reported claims.

• Proof of Benevolence: novel consensus for on-chain charitable action verification
• Zero-Knowledge Proof layer: donors prove contribution without revealing identity, amount, or recipient
• Oracle Consensus with anti-Sybil mechanism prevents gaming of the verification system
• $VELD utility token — utility-first design for governance and incentive alignment, not speculative
• Makes verified impact a financial primitive — infrastructure for trustless humanitarian finance
• Enables ESG compliance with mathematical proof rather than self-reported greenwashing claims

AI assistant that understands Indonesian capital markets in context, not just as tokens. Built by curating a custom instruction-response dataset for IDX stocks, then fine-tuning LLaMA 3.2 with LoRA using proper Llama 3 chat template formatting — not the deprecated [INST] approach most tutorials still use.

• LLaMA 3.2 3B-Instruct + LoRA (r=16, alpha=32) targeting all attention and MLP layers
• 4-bit NF4 quantization via BitsAndBytes — runs on consumer GPU under 6GB VRAM
• Dual-voice TTS: Ryan Neural (masculine, authoritative) for alarms, Libby Neural for briefings
• 5 scheduled intelligence sessions: Fajr alarm, morning market briefing, Dzuhur, Asr, Malam
• Live CoinGecko crypto prices + WeatherAPI integration built into every briefing
• Google Speech Recognition with routing to TradingView or Stockbit by voice command

Terminal-based multi-asset intelligence dashboard covering Global Macro, US Mega Tech, US High Beta, Crypto L1, Crypto Degen, Commodities, IDX Blue Chip, IDX Movers, IDX Gorengan — simultaneously, one view, fully responsive from narrow terminal to ultrawide. Not a wrapper — every metric computed from raw price data.

• 50+ metrics per asset: MA20/50/200, RSI, MACD, Stochastic, ATR, Bollinger, Z-Score, Hurst
• Kelly Criterion, Pivot Points, Fibonacci 61.8%, VWAP gap, Relative Volume, OBV direction
• Cornish-Fisher VaR, Sortino, Calmar, Omega ratio per asset
• AI alpha analysis via Ollama LLaMA 3.1 8B — quantitative hedge fund manager framing
• DuckDuckGo news scraping with site-targeted queries: Kontan, CNBC Indonesia, Bloomberg
• Quantum Score composite signal: ALPHA / VALUE / LONG / AVOID — per asset, per session

Three concurrent threads running simultaneously: audio capture → STT transcription → LLM summarization. Queue-based inter-thread communication with 10-second audio chunks. Every 15 seconds the full running transcript is summarized and follow-up questions generated — all without a single API call leaving the machine.

• Faster-Whisper medium model on CUDA float16 — real-time <2s latency per chunk
• librosa resampling: 48kHz device capture → 16kHz Whisper input format
• VAD (Voice Activity Detection) filter — ignores silence, reduces background noise
• LLaMA 3.1 8B via Ollama — key point extraction + follow-up question generation every 15s
• CustomTkinter dark GUI: live transcript panel, summary panel, suggestions panel side-by-side
• 100% local: no Whisper API, no OpenAI, no cloud — data physically stays on the machine

Executive search intelligence at VC scale — scanning LinkedIn and GitHub for founder signals, senior engineers, and stealth startups across Indonesia's tech ecosystem. Role extraction, company matching, and signal scoring all automated. Anti-garbage filter removes noise before export.

• 40+ target companies: Tier 1 Unicorns (Gojek, Tokopedia) through stealth startups
• Role hierarchy detection: Founder / Co-Founder → VP/Head → Director → Senior Engineer
• Signal scoring: 🔥 FOUNDER/STEALTH, 🚀 HIGH (Exec), ⭐ MEDIUM (Senior) per profile
• Anti-garbage filter removes job listings, ATM, promo, location noise from results
• GitHub scanner: active Indonesian repos + "stealth", "building", "founder" keyword signals
• Edge + Selenium for Google SERP with site-targeted LinkedIn operator search queries

Domain-specific LLM for Indonesian equity analysis. Key implementation insight: Llama 3.2 doesn't use [INST] format. This uses apply_chat_template() with structured message arrays — the tokenizer converts to the correct special token format the model actually expects. Most tutorials get this wrong.

• LLaMA 3.2 3B-Instruct with 4-bit NF4 quantization via BitsAndBytes — consumer GPU
• LoRA: r=16, alpha=32, dropout=0.05, all attention and MLP projection layers targeted
• Custom dataset: PE ratio interpretation, ROE analysis, RSI signal generation for IDX stocks
• Proper Llama 3.2 chat template via apply_chat_template() — not deprecated [INST] format
• float16 enforced over bf16 for Windows CUDA stability (NotImplementedError fix)
• Gradient checkpointing + enable_input_require_grads() for memory-efficient training