I'm a software developer building SiloCRM and SiloCSM for JWeis Marketing, and before that for Sharetasking Inc. Most of my recent work is LLM application engineering: a real-time voice agent on ElevenLabs ConvAI, AI-written SMS on Twilio, RAG with per-tenant knowledge-base sync, MCP tools that expose CRM data to agents, and a visual workflow automation engine.
I care about the part most LLM work skips: proving the thing actually works. That's why I built EvalLab.
Open to senior full-stack and LLM application roles, remote with US teams.
SiloPilot, a per-tenant AI agent for calls, SMS and email. It qualifies leads, books appointments, collects data and routes calls, each agent on a tenant-owned Twilio number. My part:
- the ElevenLabs ConvAI voice integration: real-time WebSocket, document store and RAG, agent provisioning
- the Twilio layer: calling, conference calls and AI-written SMS replies
- webhook ingest for transcripts, audio and conversation events
- MCP integrations and tool use for call transfer and data collection
- the AI agents management UI
The workflow automation engine. A no-code canvas (React Flow and Zustand) on a custom Fastify execution engine: 33 triggers, 20+ actions, conditional branches. It walks the graph breadth-first, one depth at a time, and logs every node, so each run can be replayed. The header above is that idea, animated. A real workflow looks like this:
flowchart LR
T([New lead]) --> A[Tag inbound]
A --> W[Wait 5 min]
W --> Q{Replied?}
Q -- no --> S[Send SMS]
Q -- yes --> C[Create task]
S --> L[(Per-node<br/>execution log)]
C --> L
Both live in a private company repo. Everything below is my own public work.
| Project | What it is | Links |
|---|---|---|
| EvalLab | LLM evaluation platform. LLM-as-judge scoring on accuracy, completeness, tone and safety; run-to-run regression checks; failure clustering; judge calibration against humans with Cohen's κ. Next.js, FastAPI, PostgreSQL. | repo · live |
| Zaxvio CRM | Multi-tenant field-service SaaS for small HVAC teams: scheduling, invoicing, customer history. Turborepo with Next.js, Fastify, Supabase and Drizzle. Built solo. | repo · live |
| Bangla sentiment fine-tuning | How much labelled data a QLoRA-tuned 1–2B model needs to match BanglaBERT, on a 6 GB laptop GPU. Reproduced the published baseline and found 7.2% train/test leakage in SentNoB. | repo |
| TabbyDeck | Turns a phone or tablet into a Stream Deck for Windows. Node host, Preact deck and editor, 399 tests. Open source. | repo · site |
| DropMate | Chat and send files to people at home over your own Wi-Fi. Windows and Android, no cloud, no account. Flutter. | repo · site · download |
| TabDeck | A feature-packed new-tab page for Chrome. My most-starred repo. | repo |
Autism Spectrum Disorder Identification from Facial Images Using Fine-Tuned Pre-trained Deep Learning Models and Explainable AI Techniques
Hossain, S. S., et al. (2025). Semarak International Journal of Applied Psychology, 5(1), 29–53. First author.
Compared 12 fine-tuned CNNs; DenseNet121 reached 90.33% accuracy and 90.00% F1. LIME and Grad-CAM show which facial regions drive each prediction.
The full stack, by layer
LLM: prompt and context engineering · agents and tool calling · MCP · RAG · structured outputs · LLM-as-judge evaluation · QLoRA fine-tuning
Interface
Service
Data and infra
Practice
How I work
- Design the API before writing it. The contract is what matters; the handler is a detail.
- Tenant isolation belongs in the database. With PostgreSQL row-level security, a missing
WHEREclause fails the query instead of leaking data. - Measure before trusting. An LLM feature without a regression check has no way to tell you it broke. That's why EvalLab exists.
- I use Claude Code and Cursor every day: directing agents, curating context, and reviewing what they write.
I'm looking for senior full-stack roles, especially LLM application work: multi-tenant SaaS, agent plumbing, workflow tooling, anywhere the engineering has a real job to do.
Outside code, it's history, storytelling and the occasional YouTube rabbit hole. Happy to swap notes on either.
