Deeptanshu Banik · Agentic AI · Automation · Orchestration

I build AI workflows that hold up after the demo.

Agentic AI and orchestration, approached as a systems problem: deterministic where it can be, a model where it has to be, a person where it matters. Backed by an analyst's habit of asking what the business actually needs.

Now: Building Nuvakin — Intelligence Systems for Education and Career
Study: B.Tech CSE (AI & Robotics), VIT Chennai, 2024–2028

See the work OASIS case study deepbanik0506@gmail.com

Fig. 1: How OASIS turns a page into a trusted record

Select any component to see what it does.

Input
Deterministic step
Model call
Stored data

Principles

Working hypotheses, with evidence

Four things I believe about building systems. Each one points to the place in the work where it was tested.

A working demo is not the same as a reliable system.
OASIS keeps 4 of its 50 configured universities live. Every other candidate is waiting on a regression gate that has not yet let one through.
The gate in OASIS
Business problems are often more interesting than the technology used to solve them.
The discount study asks a commercial question: at what point does a discount stop paying for itself? The tools are ordinary on purpose.
The discount question
AI should augment human thinking, not replace it.
In OASIS, the system drafts configuration fixes, but a person approves every one, and promotion to the live registry is a manual edit.
Human approval by design
Intelligence is knowing when to do what, using the tools at hand, and changing your mind when the argument is better.
OASIS began with no headless browser by design. The evidence changed, so I added one as a scoped fallback and then measured where it did not help.
A reversed decision

Work

Systems, prototypes and studies

Each project is labelled by maturity and by where its evidence comes from: the codebase, real use, or a dataset. Nothing here is a business outcome unless it says so, and none currently does.

In developmentBuilding Nuvakin2026 — ongoing

OASIS University intelligence engine

A pipeline that turns inconsistent official university course pages into structured, validated records, plus an explainable recommendation engine on top. Deterministic parsers do the work; a local LLM may only fill fields they left empty.

Source code is private. Architecture and evidence are shown; no company-internal information is.

858offline regression tests across 46 files, run in CICodebase
5isolated layers, with isolation enforced by static import testsCodebase
4 / 50universities live; none of the 46 candidates has passed the promotion gate yetObserved in use

Registryconfig per uni
Stored data Per-university settings live in configuration, not code. A university's status (candidate or active) is a field here.
Onboarding runsample pipeline pass
Deterministic step A new candidate gets a sample run through the real pipeline before anything is trusted.
Gap reportP0–P3 issues
Deterministic step Missing or suspect fields are classified by severity, so effort goes to what blocks correctness first.
Config suggestionhuman approves diff
Human decision The system drafts configuration changes. A person approves them; nothing is applied automatically.
Regression gatepromote/hold/reject
Deterministic step Promotion to active is a manual registry edit, made only after a PROMOTE verdict. Current state: 0 promote, 3 hold, the rest reject.
Discoverfind course pages
Deterministic step Finds candidate course pages from listing pages using clustering and scoring heuristics.
Fetchstatic first
Deterministic step Plain HTTP first. A headless browser is a scoped fallback, switched on per configuration, not the default.
Validate pagedrop non-course
Deterministic step Pages that aren't real course pages are skipped before any evidence is saved.
Extractevidence saved
Deterministic step Raw HTML and extracted text are stored as evidence, so every field can be traced back to its source page.
9 field parsersdeterministic
Deterministic step Config-driven parsers for degree, duration, fees, modules, intakes, campus, study mode, language and entry requirements.
LLM fillonly empty fields
Model call One narrowly scoped local-model call, given only the fields the parsers left empty.
Mergedeterministic wins
Deterministic step If a parser found a value, the model's answer is discarded. The LLM can add facts; it can never overwrite them.
Normaliseschema + canonical
Deterministic step Schema validation plus canonical degree names, durations, intakes, campus names and currency.
Change detectiondiff vs last run
Deterministic step Each run is compared with the stored record, so changes on a university's site show up as a report.
Structured recordsone per programme
Stored data One JSON record per programme. Duplicates are quarantined (moved, not deleted) with a manifest.
Matching engineexplainable scoring
Deterministic step Mandatory criteria plus weighted scoring, with every result explained. It never imports the ingestion code.
Document intaketranscript → profile
Deterministic step Turns an uploaded transcript into the same profile shape a manual form produces. Isolated from every other module.
APIFastAPI
Deterministic step Serves universities, courses and recommendations, with OpenAPI docs. It reuses the engine and never duplicates logic.
Review dashboardinternal
Deterministic step Internal UI for testing profiles against real records. It talks to the API over HTTP only.
End-user productnot built yet
Not built No public product runs on OASIS yet. It is shown here so the diagram doesn't overstate what exists.

Selected stepSelect a step in the diagram, or focus it and press Enter, to see what it does.

Stored data
Deterministic step
Human decision
Model call
Not built

Read the OASIS case study

The bench

Deployed

Finance Decision Council

Multi-Agent AI Financial Analysis System

Built and deployed a multi-agent financial analysis system using LangGraph, orchestrating valuation, risk, and momentum agents through parallel execution and deterministic synthesis.
Developed evidence-grounded agent outputs with structured validation, explicit disagreement reporting, evidence-gap detection, and failure handling.
Engineered a full-stack application with FastAPI backend and React frontend, integrating financial market data, interactive analysis dashboards, historical price visualizations, and JSON export.
Implemented analytical safeguards distinguishing insufficient evidence, directional signals, execution reliability, and heuristic confidence from investment advice.

Python · LangGraph · FastAPI · React · Groq · yfinance · Pydantic

Live demo Repository

StudyMay 2026

Discount & Profitability Study

Where discounting stops paying for itself

Data Public dataset: Sample Superstore (US retail)

Two earlier analyses merged into one: SQL and pandas work on a public retail dataset to find the discount level at which orders stop making money, and which categories carry the losses.

20% → 30%average order profit is positive at 20% discount and negative from 30%Dataset finding
≈20%of revenue came from loss-making order linesDataset finding
Furnitureweakest category margin; average margin turns negative at 30% discountDataset finding

Python · Pandas · SQL (SQLite) · Tableau · Seaborn

Case study Repository

StudyMay 2026

Churn Modelling on Synthetic Data

Choosing a threshold, not just a model

Data Synthetic dataset: 100,000 generated telecom customers

A class-balanced logistic regression on a synthetic churn dataset, used to explore the trade-off between catching churners and raising false alarms.

70% → 82%recall when the decision threshold drops from 0.5 to 0.4Dataset finding
0.77ROC AUC at the default threshold (accuracy 68.8%)Dataset finding
Contracttype dominates the model; month-to-month carries the highest riskDataset finding

Python · Pandas · scikit-learn

Case study Repository

StudyMay 2026

Credit-Card Customer Segmentation

Honest clusters from weak structure

Data Public dataset: Credit Card Dataset for Clustering

K-Means on 17 standardised behavioural features, with the number of clusters chosen by silhouette score and the result read cautiously.

K = 3the best silhouette of K = 2–10, at 0.247Dataset finding
8,636customers clustered after dropping rows with missing valuesDataset finding
68 / 18 / 14% split: moderate users · cash-advance heavy · high spendersDataset finding

Python · Pandas · scikit-learn · K-Means · PCA

Case study Repository

Capabilities

What I can do, and where it's proven

No self-rated skill bars. Select a capability to see which projects demonstrate it.

LLM orchestration inside deterministic systems. Models confined to the narrow jobs rules can't do, with output that can never overwrite known facts. Shown in OASIS and Finance Decision Council.

LLM orchestration inside deterministic systems
Models confined to the narrow jobs rules can't do, with output that can never overwrite known facts.
OASIS Finance Decision Council
Reliability engineering
Regression suites, CI, promotion gates, change detection and duplicate quarantine.
OASIS Finance Decision Council
Human-in-the-loop workflow design
Review queues and approval steps placed where a wrong automated action is costly.
OASIS
Service and API design
FastAPI services over isolated modules, with layer boundaries enforced by tests.
OASIS Finance Decision Council
Analytics and SQL
Turning a commercial question into queries, thresholds and a dashboard.
Discount & Profitability Study
Modelling and evaluation
Classification and clustering, read critically: thresholds, overlap, data provenance.
Churn Modelling on Synthetic Data Credit-Card Customer Segmentation Finance Decision Council

Experience

Where the work happens

Curious. Experimental. Always building.

2026 — present
Building Nuvakin
Intelligence Systems for Education and Career
- Designed and built OASIS, the extraction and recommendation engine behind the platform.
- Prototyped a local-LLM outreach qualification workflow with mandatory human review.
May — Jul 2026
Business Intelligence & Analytics Intern
Klub Ventures
Remote
- Designed and specified an executive KPI dashboard for MySpace, an internal project-management platform: the four-section hierarchy, formulas and decision logic.
- Researched the requirements and architected an MCP-based natural-language query layer intended to let executives query MySpace data through Claude.
Sep — Oct 2024
Apprentice (Data Analytics Intern)
Coratia Technologies Pvt. Ltd.
- Cleaned and processed raw datasets with Pandas and NumPy.
- Ran exploratory analysis and built visualisations in Matplotlib and Seaborn.
- Applied basic machine-learning techniques to prediction tasks.
2024 — 2028
B.Tech, Computer Science and Engineering
Vellore Institute of Technology, Chennai
Specialisation in AI & Robotics

Certifications

Issued September 2026
AI Fluency: Framework & Foundations
Anthropic

I build AI workflows that hold up after the demo.

OASIS University intelligence engine

The bench

Finance Decision Council

Discount & Profitability Study

Churn Modelling on Synthetic Data

Credit-Card Customer Segmentation

LLM orchestration inside deterministic systems

Reliability engineering

Human-in-the-loop workflow design

Service and API design

Analytics and SQL

Modelling and evaluation

Building Nuvakin

Business Intelligence & Analytics Intern

Apprentice (Data Analytics Intern)

B.Tech, Computer Science and Engineering

Certifications

AI Fluency: Framework & Foundations