Senior backend and platform architect in Kolkata, India (UTC+5:30). Building software since 2006; delivering projects through Jumacro Software Services Pvt. Ltd., established December 2014.
I build production backends for accounting, payments, document processing and EV charging: systems that need reliable integrations, explicit permissions and evidence that they work.
Portfolio: mithundas79.github.io · CV: PDF · Contact: mithun.das@jumacro.com · Book a call
Availability: Core working hours 1–11 PM IST (UTC+5:30). Client-specific overlap arranged for US and European teams.
- Paynthr — AI bookkeeping automation, 2025–2026; engagement completed, delivered through Jumacro. Built the product, backend and AWS infrastructure; in production during the engagement from June 2026. Dropbox/email intake, OCR, Claude on AWS Bedrock with tool use, prompt caching and streaming, pgvector/Titan entity resolution, per-client accounting rules and review thresholds, QuickBooks Online and Stripe. 20+ Go services on ECS, Aurora PostgreSQL with row-level security and Terraform. A four-day staging certification reconciled 267 of 275 invoices to the cent; a separate release audit documented 210 findings.
- Electroop — EV charge-point management, 2022–present, within my Localghost engagement (2018–present). Backend architecture for live chargers: OCPP 1.6/2.0.1 over WebSockets, OCPI roaming, payments, tenant APIs and mobile backends. TypeScript on AWS with C# and Python components. Localghost is the underlying engagement; Electroop is the project name.
- AVEA — shipping and logistics project through Jumacro, 2025–2026. AWS shipping-label OCR using Textract, S3, SQS, Lambda and SNS; Laravel vendor-portal fixes and a major-version upgrade; Python/FastAPI work on a direct-carrier gateway whose production launch is not claimed.
- AWS infrastructure and cost review — Terraform codification and a 24% monthly cost reduction, from $2,039 to about $1,550, with budget alarms.
Upwork history: 47 completed engagements, 2,400+ hours and 44 rated 5.0, 2015–2020.
| Repository | What it demonstrates |
|---|---|
| filing-agent | A local LLM tool loop using Ollama, application-side evidence checks, bounded retries and run receipts. Recorded session. |
| doc-intake | TypeScript, local OCR and deterministic accounting rules. Browser demo. |
| doc-intake-service | Python/FastAPI, React, Tesseract and a scikit-learn document classifier trained and evaluated on synthetic data. |
Written scope and acceptance criteria, small reviewable pull requests, staging checks and a rollback plan. I distinguish code implemented, changes deployed and behavior verified in production. I use AI coding tools openly within a review-and-verify process that I own.
Stack: Go · TypeScript/Node.js · Python/FastAPI · PHP/Laravel · C# · React · AWS (ECS/Fargate, Lambda, Aurora, S3/SQS/SNS, CloudFront, Bedrock, Textract, CodeBuild) · Terraform · Docker · PostgreSQL/RLS · pgvector/Titan embeddings · MongoDB · Neo4j · GCP · QuickBooks Online · Stripe · Dropbox · OCPP/OCPI.
My focus is backend architecture and AI application engineering. The public classifier uses synthetic data; I do not present it as production ML evaluation. Yodlee integration work was sandbox-only.