sanjay.dutta
0%
Applied AI & Automation at Shriram Finance

Sanjay Dutta.

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

I build AI systems for lending where Python computes every number and LLMs only plan and explain. Five of them are used by business teams today.

See the work Resume
CollectionIQ · query replaysample data
ask
  • LLMPlannerintent: rank branches · metric: 30+ DPD change · region: West · top 50.9s
  • PYCompiler and validatorplan checked against business ontology · 4 steps12ms
  • PYPandas engineKPIs computed on account level data0.4s
  • LLMInsight agentreads computed results only, never raw data1.1s
Branch30+ DPD Δ
Branch W-14
+2.6 pts
Branch W-03
+2.1 pts
Branch W-22
+1.8 pts
Branch W-09
+1.4 pts
Branch W-17
+1.1 pts
Three of the five sit in one cluster, and most of the rise comes from the 31 to 60 DPD bucket. Worth an early bucket review before it rolls forward.
Illustrative replay. Branch names and figures are made up.Try the live app ↗
10+
business leaders use CollectionIQ daily
100+
plain English queries answered per day
100k+
accounts across 147+ branches, 16 regions
5
internal tools adopted by business teams
Selected work

Systems in daily use.

Deterministic PythonLLM step
Flagship · portfolio intelligenceLive, used daily

CollectionIQ

Business leaders used to wait hours for an analyst to answer a portfolio question. Now they ask in plain English and get a correct, explained answer in under a minute.

Hours→<1 min
insight turnaround
10+
leaders, daily
22
section auto report
LangGraphGeminiPandasLangSmithStreamlit
Live demo Source
CollectionIQ landing screen with upload panel and dashboard sections
Landing screen. Upload the monthly extract and the dashboard, query box and reports switch on.
How a question becomes an answer
Plain English questionNBFC vocabulary
LLMLogical planneremits a declarative plan
PYCompiler and validatorontology checked
PYPandas engineevery KPI computed here
LLMInsight agentexplains computed results
Table, chart, Excel

The model never touches the arithmetic. It decides what to compute and explains what came back. Python does the rest.

Plans, not generated code

The planner outputs a constrained plan over a fixed business ontology. A compiler turns it into Pandas steps and rejects anything invalid before it runs.

One repair, then stop

If the compiler rejects a plan, the planner gets exactly one retry with the error attached. Bounded retries keep latency and cost predictable.

Fast path with zero LLM calls

Common views are served from pre-computed results. The model is only called when a question actually needs reasoning.

Traced end to end

LangSmith traces every routing decision and stage, so a wrong answer can be traced to the exact step that caused it.

Underwriting intelligenceIn progress

FinSight

Evidence-first decision support for lending. Five document engines and 30+ checks verify a borrower across documents, then AI explains the case with every claim cited.

Case documentsbank, ITR, bureau, KYC
PY5 document engines
PY30+ risk and fraud checks
LLMGrounded review and Q&A
PYCitation check
70%
less manual review
30+
deterministic checks
<30s
per case
FastAPIPostgreSQLRapidOCRGeminiReact
Private build · walkthrough on request
Document AI · credit bureauLive

AutoCAM

Loan officers copied hundreds of bureau accounts by hand. The hard part was never extraction, it was trust. So every extraction is checked against the bureau's own printed totals.

Bureau PDF or scan
PYPyMuPDF and OCR
PYValidate vs printed totals
LLMGemini, only on failure
CAM in Excel
60 min→<1 min
per borrower
3
bureau formats
PyMuPDFTesseract OCRGemini VisionOpenPyXL
Live demo Source
Document automationMost used tool

LCC Consolidator

Branch reports existed only as PDFs. Pixel-coordinate clustering reads the layout straight off the page and builds one master dataset. Same PDF, same output, every run.

Branch LCC PDFs
PYPixel-coordinate clustering
PY91 column schema
Master Excel
● Zero LLM calls. Not every problem needs one.
<0.4s
per PDF
97%
less manual effort
91
column master dataset
PyMuPDFOpenPyXLStreamlit
Live demo Source
Data preparationLive

Get Data Forge

Teams merged, split and converted files by hand or on random third-party sites. Now it's one internal app, and the merge tools auto-match renamed and truncated headers.

Excel and PDF files
PYSchema-aware header matching
Human input only on real conflicts
9
tools in one app
0
third-party sites
PandaspdfplumberOpenPyXLStreamlit
Live demo Source

Blog Writing AgentMulti-agent · LangGraph

Router decides if research is needed, Tavily research node, Pydantic section contracts, parallel writers grounded to verified citations.

Source ↗

Vehicle Insurance MLOpsEnd to end ML · AWS

Modular training pipeline to a FastAPI service on EC2, Dockerized, CI/CD with GitHub Actions and ECR, artifacts in S3. F1 score 0.87.

Source ↗

Azure Data PipelineData engineering

Medallion architecture from Bronze to Silver to Gold with Azure Data Factory, Databricks and Delta Lake.

Experience

Where the work happens.

Jul 2025 to present
Mumbai

Applied AI & Automation

Shriram Finance Limited · one of India's largest NBFCs
  • Own portfolio analytics across 147+ branches and 16 regions, tracking delinquency, SMA and NPA in Power BI
  • Built and shipped 5 internal AI and automation tools now used by business teams, from CollectionIQ to LCC Consolidator
  • Automated recurring MIS reporting with Python, VBA and Power Query, cutting effort by 90%
  • Turn the questions leaders ask every week into self-serve systems, so answers stop waiting on an analyst
2021 to 2025
Prayagraj

B.Tech

MNNIT Allahabad · CGPA 8.02
Approach

Deterministic first.
AI where reasoning pays.

The pattern behind every project above. The diagrams show it: most steps are plain Python, and the LLM steps are few and deliberate.

Start with the bottleneck

Understand the workflow, find the recurring delay, and only then choose the architecture. Most bottlenecks are information reaching the wrong person too slowly.

Correctness before intelligence

Business numbers are computed deterministically and validated. AI plans and explains on top of results that are already right.

Measure the friction removed

A project counts when people stop double-checking it. Every tool here is judged by time saved and daily use, not by technique.

If the answer is knowable from the data, it belongs in a table, not a prompt.
A confident wrong answer is worse than a slow right one.
Most "AI projects" are deterministic software with a small reasoning layer on top.
The hardest part of automation is earning enough trust that people stop double-checking.
Skills

What I build with.

Applied AI

LangGraphAgentic AILLM IntegrationGeminiRAGContext EngineeringMulti-agent OrchestrationLangChain

AI Engineering

Document IntelligenceStructured OutputsFastAPIPydanticEvaluationLangSmithPyMuPDF · OCR

Data and ML

PythonPandasSQLPostgreSQLPower BIExcel · VBA · Power QueryScikit-learnDeep Learning

Deployment

DockerGitHub Actions · CI/CDAWS EC2 · S3 · ECRRailwayRenderStreamlitReact · TypeScript
Highlighted: used in production at work. Domain: underwriting, credit risk, collections, portfolio analytics.
Now building

In progress.

Portfolio intelligence

CollectionIQ v2

Wider question coverage and a sharper 22 section reporting pipeline.

Underwriting AI

Retrieval for borrower files

Advanced RAG so case Q&A in FinSight answers only from cited evidence.

AI reliability

Evals for agent pipelines

Test sets and tracing so every prompt or model change is measured, not guessed.

Lending operations

More workflows, same pattern

Extending automation across collections and underwriting workflows.

GitHub

Shipping, week by week.

github.com/Sanjay-00 View profile ↗
GitHub contribution graph for Sanjay-00
Contact

Hiring for Applied AI?

I'm open to Applied AI Engineer and AI Automation roles, especially in lending and fintech. Based in Mumbai. The fastest way to reach me is email.

sanjaydutta2830@gmail.com↗ View resume↗ LinkedIn↗
↑