● AI AGENT ENGINEER · FULL-STACK DEPTH
I build AI agents
that do real work.
Voice tutors, banking automations, research crews — shipped and measured. I build the agent loop and the whole production system under it.
AI agents & voice
Voice tutors, research crews, and browser agents that actually ship — agent loops, MCP tools, and sub-second voice pipelines.
6 agent systems shipped in 2026
Founding-engineer depth
I take products from nothing to launched. As Dynt's founding engineer I wrote ~82% of the codebase — open banking to card issuing.
~82% of a fintech, solo
Real-time & data UIs
DAG workflow viz, a C++/WASM data grid, offline-first POS, WebRTC voice — the hard frontend, not CRUD.
React 19 · WASM · WebRTC
About
I build AI agents that do real work, not demos that fall over when you click twice. The proof is the boring stuff underneath: at the fintech Dynt I wrote about 82% of the commits, including a webhook queue that stops banks from breaking each other when events arrive out of order.
The fun stuff sits on top. No off-the-shelf model speaks colloquial Bangla, so I trained one from a single YouTube video on a rented RTX 3090. My tutor's voice stalled because Bangla sentences end in a danda, not a period, so I taught the tokenizer the difference and got seconds back. I work from the model loop down to the Postgres view. I'm in Dhaka, and I want to teach Bangladeshi students with this.
01 — Selected Work
Agents first. Foundations underneath.
Thirteen projects across agents, fintech, real-time systems, and data. Confidential work is shown as a case study — what I built and the impact, without proprietary code.
02 — Experience
Five years, four startups, zero handoffs.
AI Agent Systems · Independent
2026 — present- Built demo-agent and Mirars end to end, and architected Mr. Topper — a real-time Bangla voice tutor — on a Pipecat voice pipeline with a NestJS backend.
- Trained a Bengali text-to-speech model from scratch — YouTube scrape, Demucs isolation, Whisper labeling, XTTS-v2 fine-tune — and served it on a $0.22/hr RTX 3090.
- Shipped agents on two transports from one tool registry (chat + MCP), so the same agent installs into Claude Code or Cursor with zero client config.
Lead Frontend Engineer · Datatailr
2024 — 2026- Owned the platform console end to end: 412 of 543 frontend commits (76%), ~37,650 lines of TypeScript across 14 routed pages.
- Cut a fleet-scale dashboard's GraphQL N+1 fan-out to one indexed read by writing 442 lines of Postgres analytics views exposed through pg_graphql (issue #1297).
- Built the DAG workflow viewer on xyflow + dagre with viewport-culled rendering, and wired a dual codegen pipeline (GraphQL Codegen + orval) into the build.
Founding Software Engineer · Dynt
Jun 2023 — Apr 2024 (full-time), then ongoing- Wrote ~82% of commits across three repos as founding engineer: 2,913 of 3,559, and roughly 61,000 lines of TypeScript.
- Built the open-banking sync layer (Swan, Yapily, GoCardless, CSV → one model) and solved out-of-order Swan webhooks with a typed, dependency-ordered EventQueue.
- Shipped Stripe card issuing with Zod-validated KYC, a json-rules-engine automation system with a full audit trail, and a CrewAI bookkeeping crew (May 2024).
Full-Stack Engineer · Neoke
2024 — 2025- Built Neoke's traveler ID-verification web flow: webcam document and selfie capture, OCR/MRZ result display, and a face-match confidence dial.
- Integrated IATA Timatic's travel-requirements API (OAuth2 client-credentials) behind an Express/Bun proxy with Zod-validated schemas.
- Built the marketing site's blog and FAQ on a Strapi CMS backend with a React front end.
Sole Engineer · Possier
2022 — 2023- Sole engineer on a restaurant-POS vendor's products serving 200+ clients across a consumer social platform, the legacy ERP, and a ground-up POS rebuild.
- Built Fuvii, a social food-video platform (Express + TS + Prisma, 18 models, 28 endpoints) with an engagement-ranked feed and direct-to-S3 presigned uploads.
- Architected the ground-up multi-tenant POS rebuild (Turborepo + NestJS + Drizzle, 26 tables, offline-first sync, atomic multi-tender checkout).
03 — Stack
Tools I bet products on.
- LangGraph
- LangChain
- CrewAI
- Model Context Protocol (MCP)
- LiveKit Agents
- Pipecat
- Gemini
- OpenAI
- Groq
- Deepgram STT
- Cartesia / ElevenLabs TTS
- pgvector
- PyTorch
- XTTS-v2 / Fish Speech
- Stagehand
- Playwright
- TypeScript
- Python
- NestJS
- Hono
- Fastify
- Express
- tRPC
- Prisma
- Drizzle
- PostgreSQL
- Supabase
- Redis / BullMQ
- GraphQL / pg_graphql
- Zod / Pydantic
- React 18 / 19
- Next.js
- Vite
- TanStack Query
- Apollo Client
- Tailwind
- xyflow + dagre
- Monaco
- React Native
- Tauri
- Canvas / WASM bridges
- Browserbase
- ffmpeg
- Docker
- Fly.io
- Vercel
- AWS S3
- Turborepo / pnpm
- GraphQL Codegen + orval
- WebRTC / Socket.IO
- RTX 3090 model serving
// from thermal receipt printers to voice agents — the stack just keeps getting weirder
04 — Signal
What people say.
A quote on founding-engineer ownership and trust.
A quote on lead-frontend ownership and the fleet-scale perf fix.
A quote on social proof on the merged PR.
A quote on the independent 2026 work shipping and working.
05 — Contact
Let's build an agent
that ships.
Open to AI agent engineering work — contract or full-time. The hard problems are the fun ones.