I'm a full-stack engineer at Sharetasking Inc., building SiloCRM with a team of five. Most of my recent work is LLM application engineering: a real-time voice agent on ElevenLabs ConvAI, AI-generated SMS on Twilio, RAG with per-tenant knowledge-base sync — and a visual workflow automation engine I designed and built end-to-end, solo.
I care about the part most LLM work skips: proving the thing actually works. That's why I wrote EvalLab.
Khulna, Bangladesh (UTC+6) · remote with US teams · open to senior full-stack roles
The engine walks a trigger graph breadth-first, one depth at a time, and writes a log line per node. Every run is replayable in a read-only viewer. This is the shape of a real workflow a user builds on the canvas:
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
React Flow canvas + Zustand on the front, a custom Fastify execution engine behind it. 33 trigger types, 20+ actions across the CRM lifecycle.
What I'm shipping right now
SiloPilot — a per-tenant, multi-channel AI agent for lead qualification, appointment booking, data collection and call routing. Each agent is bound to a tenant-owned Twilio number. I built:
- the ElevenLabs ConvAI voice integration — real-time WebSocket, document store / RAG, agent provisioning
- the Twilio SMS layer with AI-generated replies
- webhook ingest for transcription, audio and conversation events
- RAG pipelines with per-tenant knowledge-base sync
- tool-use for call transfer and data collection
- the AI agents management UI
Workflow Automation Engine — owned end-to-end, solo. Visual no-code builder, custom execution engine, breadth-first graph traversal, per-node logs, replayable run viewer.
Both are company products behind a private repo, not side projects. The public repos below are my own work.
Selected work
| Project | What it is | |
|---|---|---|
| EvalLab | LLM evaluation harness — LLM-as-judge scoring across four dimensions, failure clustering, judge calibration with Cohen's κ and a 5×5 confusion matrix. Python/FastAPI + Next.js. | repo · live |
| Zaxvio CRM | Multi-tenant field-service SaaS for HVAC contractors. Scheduling, quoting, invoicing, customer history. Built solo, end-to-end. | repo · live |
| SiloCRM | Multi-tenant AI-powered CRM. SiloPilot and the workflow engine live here. | live |
| TabDeck | A Chrome new-tab page. Open source. | repo |
| Counterfoil | SaaS platform — frontend contribution at Ternary Solutions. | live |
The stack, by layer
LLM — Prompt engineering · Agent design · RAG · Tool-use / function-calling · Streaming inference · LLM-as-judge evaluation · Prompt versioning
Interface
Service
Data & infra
Multi-tenant isolation is enforced at the database with PostgreSQL row-level security, not in application code.
Practice
How I work
- REST APIs get designed before they get implemented. The contract is the artifact; the handler is a detail.
- Tenant isolation belongs in the database. Row-level security means a missed
WHEREclause is a failed query, not a data leak. - Evals before vibes. If an LLM feature ships without a way to measure regression, it ships without a way to know it broke. That opinion is why EvalLab exists.
- Daily Claude Code and Cursor user — directing agents, curating context, and reviewing what they produce across the monorepo.
I'm open to senior full-stack roles, particularly LLM application work. Multi-tenant SaaS, agent plumbing, workflow tooling — anything where the engineering has a real job to do.
Outside of code it's history, storytelling, and the occasional YouTube rabbit hole. Happy to swap notes on either.
