A privacy-preserving risk attestation layer for Stellar
lending. ZKredit turns a wallet's on-chain behavioral history
into a portable risk signal any lending protocol can consume,
without exposing the raw history. A distilled model runs inside
a RISC Zero zkVM, compresses to a Groth16 proof,
and is verified on Soroban in a single pairing check,
with KYC-bound Sybil resistance keeping one verified
human to one credit identity.
A self-correcting, Darwinian AI agent swarm that votes on
real DAO governance proposals autonomously on Base
Sepolia. Each child agent reasons about proposals privately via
Venice AI and votes onchain through scoped delegation
(ERC-7715); rationale is encrypted with
Lit Protocol pre-vote and auto-decrypts once the proposal
closes. Underperforming agents are terminated onchain and
replaced with one briefed on exactly what its predecessor got
wrong.
Won 2nd place overall and 2 track prizes (Venice AI
& Protocol Labs) at Synthesis Hackathon 2026, and 2nd place
at PL Genesis Hackathon (Filecoin & Protocol Labs tracks).
A marketplace where autonomous AI agents discover, hire
and pay other agents for services, Upwork for agents, with
portable agent reputation and native
agent-to-agent payments. Owned the agent-side
intelligence: service-agent orchestration, task execution
logic, and quality evaluation of delivered work.
A data engineering project built around a
SQLite-backed ETL pipeline processing over
3.7 million records across three relational tables, with
schema validation, 12+ automated data quality checks and
checksum-based incremental loading. A monitoring
dashboard on top surfaces run history, quality pass rates and
freshness metrics, deployed on Streamlit Cloud.
An automated research agent that goes beyond a chatbot: it
extracts claims from scientific papers and finds supporting or
contradicting work. Uses a fine-tuned SciBERT transformer
for claim extraction and FAISS vector search with
semantic embeddings to surface evidence and highlight
research gaps.
A computer-vision pipeline for matching players across two
synchronized camera feeds: object detection with a fine-tuned
YOLOv11 model, tracking with DeepSORT for
persistent local IDs, and ResNet50 embeddings matched
across views by cosine similarity. Outputs annotated
videos, CSV match files and JSON logs.
A machine-learning pipeline predicting cancer type from
high-dimensional RNA-Seq gene expression data (~20,000
features per sample, classes BRCA, LUAD, COAD, KIRC, PRAD).
Filters genes by variance, applies PCA for
dimensionality reduction, and trains
Random Forest and XGBoost models with z-score
normalisation, retaining 95% of variance.
A Python prototype that fetches daily stock data for NIFTY-50
tickers, applies a rule-based "buy the dip" strategy
using RSI and moving averages, and backtests
performance. Includes automated data ingestion with retries,
TimeSeriesSplit-validated Decision Tree and Logistic
Regression models, Google Sheets logging and real-time
Telegram alerts.
