Sailab Banik

An AI document processing platform where business teams create their own extraction models, annotate the fields they need, and push the results straight into their workflows. I led the frontend and built role- and group-based access control from the ground up so teams with strict data rules could adopt it.

• React
• Redux Toolkit
• FastAPI
• PostgreSQL
• Azure DevOps

Business teams onboarded

5+

Less manual document review

60%

Operational hours saved each quarter

2,000+

A support chat built to take Redis Iris apart: every request returns an answer and, beside it, the record of how that answer was made — which memories and records reached the model, what was left out and why, and what each step cost. It exists to put a number on what a semantic cache is worth, so a question asked again in other words comes back from LangCache with the model never called.

• Python
• FastAPI
• Redis Iris
• Next.js
• TypeScript
• Docker

A rephrased question answered from cache, against 4.65s through the model

917ms

Tokens billed on that answer, down from 576

0

Similarity the rephrasing matched at, against a 0.92 threshold

1.00

Source

A job queue where FastAPI accepts work, Redis hands it to a background worker, and PostgreSQL records every job's state so nothing is lost across restarts. A Next.js dashboard shows each job moving through the queue, including failed attempts that retry with exponential backoff.

• Python
• FastAPI
• Redis
• PostgreSQL
• Next.js
• Docker

Attempts per job, backing off 2s then 4s

3

Jobs per client every 10 seconds before a 429

10

Waiting jobs before the API sheds load

100

Source

A pipeline that turns unstructured PDFs into JSON matching a schema the user defines, running entirely on local infrastructure so nothing sensitive leaves the environment. Text and tables live in separate vector stores, chunked along the document structure and reranked by a cross-encoder.

• Python
• LangChain
• Qwen3
• Chroma
• FastAPI

Better retrieval precision than a single-store baseline

+40%

Valid JSON responses, up from 75%

100%

I change how the team ships, not just what I ship

Release cycle time came down 40% and production hotfixes 35% because I went after the pipeline the whole team works through rather than my own throughput. The frontend test suites went in for the same reason. Work counts when it changes how everyone ships, not how fast one person does.

I lead by making the reasoning visible

Leading the frontend of PrimeAI is mostly not writing frontend. It is making a decision legible enough that the next person can pick it up, argue with it, or replace it without asking me first. Access control reached the teams with the strictest data rules because they could see exactly what it would and would not allow, not because I told them it was safe.

I build systems that are measurable, not just functional

ContextDesk Lab exists because I did not want to take a semantic cache's word for it. Every request returns an execution record, a miss reports its score, and no component is ever shown as working when it was not. I build the same way at work: a claim is worth whatever the measurement behind it is worth.