sanjay.dutta
0%
// Applied AI Engineer · Analytics · Builder

Sanjay
Dutta.

I don't start with AI.
I start with the bottleneck.
Then I choose the architecture.

Business users waited hours for portfolio insights. Underwriters spent nearly an hour processing bureau reports. Operations teams manually merged reports every day. Those weren't AI problems. They were information bottlenecks. I build systems that remove them.

Used Daily by Business Teams Portfolio Analytics · Shriram Finance 📈 Lending & Credit Risk
View Projects LinkedIn ↗ GitHub ↗
100+ Queries a Day · CollectionIQ
70k+ Account Portfolio
Hours → Under 1 Minute
147+ Branches · 16 Regions
01 · Experience

Where I have worked.

Aug 2024 · Present
Mumbai
Portfolio Analytics and Applied AI
Shriram Finance Limited · Leading NBFC
  • Own portfolio analytics across 147+ branches and 16 regions covering 10,000+ monthly loan accounts
  • CollectionIQ — portfolio insight turnaround down from hours to under one minute, used daily by 10+ collection leaders and regional heads
  • AutoCAM — CIBIL bureau extraction across CRIF and TransUnion formats, self-validated, 60 minutes down to under 1 minute per case
  • FinSight — hybrid deterministic and AI underwriting workflow, 15 rule-based validations plus Gemini risk narratives, 70% less review effort
  • LCC Consolidator — pixel-coordinate clustering with dynamic column handling, 97% less manual effort · Get Data Forge — repetitive data preparation eliminated
  • Recurring MIS reporting effort down 90% through VBA and Power Query automation
  • Delinquency, SMA, and NPA tracked across 16 regions through Power BI dashboards
02 · About

Who I am.

Working in lending taught me that businesses rarely suffer from a lack of data. They suffer because useful information reaches decision makers too slowly. Most of what I build exists to remove that delay.

B.Tech graduate from MNNIT Allahabad (2025, CGPA 8.02), working in Portfolio Analytics and Applied AI at Shriram Finance Limited, covering 147+ branches and 16 regions across India.

I observe workflows, find the recurring friction, and build reliable systems to eliminate it. Sometimes the right answer is Python automation. Sometimes SQL. Sometimes OCR. Sometimes AI. Sometimes a hybrid.

Technology follows the problem.
Not the other way around.

In lending, incorrect numbers matter more than slow numbers. That is why I keep business calculations deterministic and use AI only for reasoning and interpretation, never for the numbers themselves.

0hrs
to under 1 min · CollectionIQ
0min
to under 1 min · AutoCAM
0%
faster underwriting review · FinSight
0%
MIS reporting automated
0
internal tools adopted by business teams
0+
branches · 16 regions
03 · Projects

What I have built.

Featured Project
Lead Project · Used DailyBuilt & Live
CollectionIQ
The Problem
  • Business leaders waited hours for analysts to answer portfolio questions
  • The data already existed; the bottleneck was access
The Solution
  • Ask questions in plain English, get answers in under a minute
  • AI understands intent and explains results
  • Every KPI, ranking, and number computed deterministically in Python, never by the model
100+
Queries / Day
70k+
Accounts
<1min
Insight Time
147+
Branches
Why AI never computes the KPIs
Business users don't need creative numbers. They need correct numbers they can act on. Every KPI is computed in Python against the raw data. AI only reasons over outputs that are already correct, it never touches the arithmetic.
Landing Page
Landing
LangGraph Gemini 2.5 Flash-Lite Pandas Streamlit SMTP Plotly LangSmith
Live Demo GitHub
Architecture · AI Query Pipeline (LangGraph)
Plain English Query
User input in NBFC domain language
↓
Logical Planner Agent
Gemini 2.5 Flash-Lite · reads registry vocabulary · emits declarative intent · clarifies when ambiguous
↓
Compiler and Validator
Deterministic · lowers intent into pandas step-plan · validates schema · one LLM repair on failure
↓
Data Executor
Pandas · step-plan engine · KPI computation · priority framework · rankings
↓
Insight Generator Agent
Gemini 2.5 Flash-Lite · domain-aware observations over computed results
↓
Result
Filtered loan table · ranked aggregation · single stat · Excel download
Orchestrated with LangGraph · Fast-path view serves pre-computed results without any LLM call · Stateless per query
LLMs never touch raw calculations. All business metrics computed deterministically in Pandas.
Other Projects
Applied AI Systems
AI Underwriting · LiveIn Progress
FinSight
The Problem
  • Underwriters spend significant time validating borrower information
  • Reviewing bureau reports and preparing credit assessments eats the day
The Solution
  • AI only interprets qualitative remarks branch managers write
  • Deterministic logic handles every numerical flag
70%
Review Cut
15
Checks
<10s
Per Case
Gemini 2.0 Flash LangChain PyMuPDF Python Streamlit
Live Demo GitHub
Multi-agent Pipeline · LangGraphComplete
Blog Writing Agent
  • Multi-agent pipeline mirroring how a technical writer works
  • Router classifies topic, decides if research is needed
  • Research node gathers Tavily results, dedupes by URL
  • Orchestrator plans each section with Pydantic-enforced contracts
  • Workers write sections in parallel, grounded to verified citations
  • Merge node restores order and assembles the final document
⚡ Agentic routing · fan-out parallelism · schema-enforced section contracts
LangGraph Gemini 2.0 Flash Tavily Pydantic v2 Streamlit
GitHub
Automation & Data Systems
Document AI · Live · Multi-formatBuilt & Live
AutoCAM
The Problem
  • Loan officers manually extracted hundreds of accounts from bureau reports
  • The hard part was never extraction. It was trust
The Solution
  • Validates every extraction against the bureau's own printed totals
  • Gemini only invoked when validation fails, never as the primary path
60min
Before
<1min
Now
3
Bureau Formats
PDF or Scan→OCR→Deterministic Extraction→Validation→AI only if required→Excel
Tesseract OCR Gemini Vision PyMuPDF NumPy OpenPyXL ThreadPoolExecutor
Live Demo GitHub
Data Preparation · Live · 9 ToolsBuilt & Live
Get Data Forge
The Problem
  • People handled basic Excel and PDF tasks manually, one file at a time
  • Or bounced between random third-party websites just to merge, split, or convert a file
The Solution
  • Merge and split tools are schema-aware, auto-matching renamed and truncated headers
  • Only asks for human input where there's a genuine conflict
9
Tools
1
App
0
Third-Party Sites
Pandas Python OpenPyXL pdfplumber Streamlit
Live Demo GitHub
Custom Extraction · LiveBuilt & Live★Most Used Tool
LCC Consolidator
The Problem
  • The information already existed. Getting it into a usable form was the real problem
The Solution
  • Pixel-coordinate clustering parses layout directly off the page, no brittle table library
  • Deterministic and repeatable, same PDF, same output, every run
<30s
Full Consolidation
97%
Effort Cut
Dynamic
Column Handling
Pixel-coordinate Clustering PyMuPDF OpenPyXL Streamlit
Live Demo GitHub
Infrastructure & MLOps
MLOps · Data Engineering · CloudComplete
Vehicle Insurance Prediction + Azure DE Pipeline
  • MLOps pipeline: MongoDB, AWS S3, ECR, EC2, GitHub Actions CI/CD
  • Azure DE Pipeline: Medallion architecture, Bronze to Silver to Gold
0.87
F1 Score
2
Full Pipelines
3
Bronze/Silver/Gold
Docker Databricks FastAPI AWS Azure Data Factory Delta Lake
GitHub
04 · Philosophy

The pattern behind the work above.

01
Understand the workflow
↓
02
Find the recurring bottleneck
↓
03
Choose the architecture
↓
04
Build deterministic components wherever possible
↓
05
Introduce AI only where reasoning adds value
↓
06
Measure business impact
AI is not the default
Start with the bottleneck. Choose the architecture after understanding the problem, not before.
Correctness before intelligence
Business calculations stay deterministic. AI explains, it does not compute.
Measure business impact
Every project should remove measurable operational friction, not just demonstrate a technique.
05 · Who This Is For

Built for teams running on manual work.

📊

Portfolio & Risk Teams

  • Tracking DPD, SMA, NPA across branches
  • Need self-serve intelligence
  • Tired of waiting on analysts
Portfolio Analytics
📄

Credit & Underwriting Teams

  • Reviewing bureau reports by hand
  • Need validated extraction
  • Correctness cannot be sacrificed
Document AI
⚙️

Ops & Data Teams

  • Manually reconciling headers
  • Merging files every cycle
  • Friction is the real cost
Workflow Automation
🤖

Applied AI Teams

  • Building agentic systems for production
  • Correctness cannot be probabilistic
  • AI reasons, logic decides
Applied AI
06 · Notes

Working notes.

Why AI shouldn't calculate KPIs. It should reason over numbers someone else already trusts.
Validation matters more than extraction. A confident wrong answer is worse than a slow right one.
If the answer is knowable from the data, it belongs in a table, not a prompt.
Most "AI projects" I've built are mostly deterministic software with a small reasoning layer on top.
Most bottlenecks I've run into aren't AI problems. They're information reaching the wrong person too slowly.
The hardest part of automating a workflow is rarely the code. It's earning enough trust that people stop double-checking it.
07 · Skills

What I work with.

Core Proficient Familiar
Reasoning
LangGraphUsed inCollectionIQ · Blog Writing Agent
Gemini 2.5 Flash-LiteUsed inCollectionIQ query and insight agents
LangChainUsed inAutoCAM · FinSight
Multi-agent OrchestrationUsed inCollectionIQ 4-agent pipeline
RAG · TavilyUsed inBlog Writing Agent research node
Prompt EngineeringUsed inEvery agent across all projects
Data
Python · Pandas · NumPyUsed inEvery project I've built
SQLUsed inPortfolio analytics at Shriram Finance
Power BIUsed inDelinquency and NPA dashboards
PyMuPDF · Tesseract OCRUsed inAutoCAM · LCC Consolidator
Excel · VBA · Power QueryUsed inMIS reporting automation
Classification · Regression · Feature EngineeringUsed inVehicle Insurance Prediction, F1 0.87
Deployment
FastAPI · DockerUsed inVehicle Insurance MLOps pipeline
AWSUsed inS3 model registry, ECR, EC2
AzureUsed inAzure DE Pipeline (Medallion architecture)
Databricks · Delta LakeUsed inAzure DE Pipeline (Medallion architecture)
GitHub Actions · CI/CDUsed inMLOps pipeline deployment
LangSmith · Claude CodeUsed inQuery tracing and debugging
08 · Working On

Currently working on.

🔄
PORTFOLIO INTELLIGENCE

CollectionIQ v2

Next iteration of the query and reporting pipeline.

2026 · In Progress
📑
UNDERWRITING AI

Enterprise Underwriting Intelligence

Advanced RAG for grounded borrower Q&A.

2026 · In Progress
🧪
AI RELIABILITY

AI Evaluation, Reliability & System Design

Evaluation frameworks and reliability patterns for production agents.

2026 · In Progress
⚙️
WORKFLOW AUTOMATION

Lending Operations Automation

Extending automation across collections and underwriting workflows.

2026 · In Progress
09 · Contact

Let's build something.

If you are solving a difficult operational problem and wondering whether AI is actually the right answer, I would love to talk.

sanjaydutta2830@gmail.com LinkedIn ↗ GitHub ↗