I build AI products end to end, from model pipelines to the interfaces people use.
A CRM for SDR and sales teams, with an MCP connector so AI assistants can work alongside the people using it.
Contacts and lead pipeline, LinkedIn outreach, opportunities, daily reports, calendar, dashboards, and notifications across in-app, email, and web push. The browser talks to Supabase directly behind row-level security, while a Hono API on Cloudflare Workers or Render handles privileged work, scheduled jobs, and an OAuth-secured MCP connector for AI clients.
An agentic financial intelligence platform built on LLM orchestration and retrieval-augmented generation.
I built an agentic financial intelligence platform combining LLM orchestration, document ingestion, semantic retrieval, conversational memory, and RAG. The FastAPI backend adds ESG scoring, multi-agent swarm orchestration, and a trust graph that keeps an audit trail of every agent's work.
An AI text refinement platform with real-time streaming, multi-pass processing, and side-by-side comparisons.
I engineered a full-stack AI refinement platform with real-time SSE streaming, multi-pass processing, batch execution, and interactive text comparisons. The Next.js frontend covers live refinement, diffing, batch results, and analytics on top of a FastAPI backend.
Turns published research into factual, verified LinkedIn posts, straight from a researcher's publication record.
A privacy-first platform for researchers: connect a research identity once through ORCID, OpenAlex, arXiv, PubMed, or Crossref, and it continuously turns new publications into authentic LinkedIn content that stays faithful to the source.
AI-powered B2B lead intelligence: automated web research, decision-maker discovery, and structured data extraction.
An in-house SmartLogics product. I built an AI-driven lead enrichment system combining automated web research, decision-maker discovery, structured data extraction, and cost-aware API routing across model providers.
Lead generation, enrichment, scoring, drafting, and outreach in one staged pipeline.
LeadForge turns raw business search results into qualified leads. Each stage, from prospect discovery and company enrichment to scoring, email drafting, and campaign preparation, runs and can be tested on its own.
An identity-verified soulbound NFT platform linking KYC verification to non-transferable tokens.
I contributed to a decentralized identity platform connecting KYC verification, non-transferable NFT minting, and authenticated blockchain transactions.
Fixed Headroom context compression for clients that run their own tool loop: retrieved content was being re-compressed to the same hash and lost, so agents looped or answered from an empty skeleton.
Fixed missing Tagalog translations in the Validator component on the 6.4 branch.
A change to the HTTP module so buffered outgoing data is counted in bytes rather than string length, letting backpressure apply when the socket buffer is actually full.
Reported and fixed an unauthenticated loopback webhook (CWE-306) that let any local process forge interaction responses; the maintainers shipped the fix in #4013, crediting my PR. I also have fixes in review for WhatsApp sender identity across channels and cross-session task visibility.
A Helm chart fix so Prometheus stops scraping the headless service when the ServiceMonitor is enabled.
Battery status for every connected pair of AirPods in the Linux tray, and an Android option to duck volume for Conversational Awareness only when both AirPods are worn.
Every pull request I have opened on other projects is in my GitHub contribution history (opens in a new tab). My own projects, from agents and automation to coursework, are on my GitHub profile (opens in a new tab).
My projects sit where AI meets product: LLM orchestration and retrieval, computer vision and OCR, and the full-stack platforms that put them in front of people. Many started as client work at SmartLogics (opens in a new tab) or as in-house products like NitLead.
I care about shipping systems that hold up in production: measurable latency, honest evaluation, and interfaces people actually use.