Jatin.Pal Open to AI/ML roles
 IST Resume

AI/ML ENGINEER · GRADUATING 2027

Jatin Pal.
I build models that ship, and measure them honestly.

Final-year B.Tech CSE (AI) student at Bennett University. I fine-tune transformers, build retrieval and recommendation systems, and wire LLMs into apps people can actually use.

See my work Resume GitHub LinkedIn
88.4%LegalBERT accuracy on unseen contracts
2live demos you can try
510lawyer-annotated contracts used

01 · PROJECTS

Things I've built

Every number here comes from the code in the linked repository, and three of these run live. Drag, swipe or use the arrows.

01LIVE · LANGGRAPH AGENT

FinReport Agent

Answers questions about company annual reports, citing the pages it used. It plans its searches, checks it found the right passages, does every calculation with a calculator tool and verifies its numbers against the report.

70%
numerical questions right (plain RAG: 50%)
66%
with a 4B model on my laptop GPU (RAG: 36%)
145
held-out FinanceBench questions, p < 0.01
  • Hybrid search (bge-small + BM25) over 84 public filings, with full-page context so tables are never cut in half.
  • My first 30-question run overstated the gain for the large model. On all 145 it helps on numerical questions but not qualitative ones, so simple lookups are now routed to plain RAG.
  • Also deployed as a streaming FastAPI service in Docker on Azure Container Apps (metrics, API keys), and as an MCP server that Claude Desktop or other agents can use.
  • LangGraph
  • LangChain
  • RAG
  • ChromaDB
  • FastAPI
  • Docker
  • Azure
  • MCP
  • Groq
  • Ollama
Try the live demo Code Evaluation API on Azure
finreport-agent.streamlit.app
The FinReport app answering how much 3M's capital expenditure changed from FY2017 to FY2018: the agent's search, grading and calculator steps, then the answer, $204 million or 14.86%, citing page 46.
02LIVE · FINE-TUNED LEGALBERT

Legal Contract Auditor

Splits a contract into clauses, classifies each one into 19 clause types with a fine-tuned LegalBERT model, flags risky wording and shows which words drove each prediction.

88.4%
accuracy on held-out contracts
0.864
macro F1 (baseline 0.811)
76
test contracts never seen in training
  • 9,149 clauses from the lawyer annotations in CUAD (510 contracts), split by contract so the test set is truly unseen.
  • My first version claimed 96.8%, but its labels came from keyword rules that alone scored 95.8%. I rebuilt the labels from the real annotations and documented why.
  • PyTorch
  • Transformers
  • LegalBERT
  • Scikit-learn
  • Docker
Try the live demo Code Model
jatin-legal-auditor.streamlit.app
The Legal Auditor app showing two clauses from a sample contract classified as Exclusivity and License Grant, with the most influential words highlighted.
03LIVE · SMART INDIA HACKATHON

InternAI

An internship recommendation system built for Smart India Hackathon, where I led the team.

6,500+
internship listings searched
3
LLM features: skill gaps, resume feedback, roadmap
  • Content-based recommender with TF-IDF and cosine similarity, filtered by location, stipend and duration. It works without any API key.
  • Gemini features run on each visitor's own key, kept per session and never stored.
  • Scikit-learn
  • Gemini API
  • Ollama
  • Streamlit
Try the live demo Code
innovace.streamlit.app
InternAI showing 10 matching internships, a 52% average match score and the top machine learning internship recommendations.
04EVOLUTIONARY SEARCH

Genetic Algorithm Feature Selection

Finds a small, strong feature subset: a mutual-information filter first, then a genetic algorithm whose fitness trades accuracy against the number of features.

30 → 5
features kept on Breast Cancer Wisconsin
94.7%
5-fold cross-validated accuracy
3
models compared: LR, random forest, SVM
  • Tournament selection, crossover and adaptive mutation, with fitness evaluated in parallel with joblib.
  • Scikit-learn
  • joblib
  • Streamlit
Code
genetic-feature-selection
The genetic algorithm feature selection app with a dataset preview and a target column selector.
05RULE ENGINE · TESTED IN CI

IT Support Agent

Resolves internal IT requests with a rule-based 5-stage pipeline, so every decision cites a written policy and can be audited.

17
Pytest tests, run on every push
15/15
scripted requests decided as expected
10
action tools, from unlocks to laptop dispatch
  • Parse, retrieve policy and past tickets, decide, act, respond, with state passed between stages as Pydantic models and every step logged.
  • Python
  • Pydantic
  • Pytest
  • GitHub Actions
Code
it-support-agent
The IT Support Agent tracing a laptop replacement request: parsed details, the matched policy KB-03, a past ticket, and the approval with its reasoning.

02 · SKILLS

What I work with

Python first, with the ML and LLM tooling around it.

Machine learning

  • PyTorch
  • TensorFlow
  • Keras
  • Scikit-learn
  • CNNs
  • GANs
  • VAEs
  • RL (DQN)

Generative AI & NLP

  • Transformers
  • BERT
  • LegalBERT
  • DistilBERT
  • LangChain
  • LangGraph
  • RAG
  • ChromaDB
  • Gemini API
  • Groq
  • Ollama
  • Hugging Face

Computer vision

  • OpenCV
  • YOLOv8
  • U-Net
  • Segmentation

Languages

  • Python
  • C++
  • Java
  • SQL

Backend & data

  • FastAPI
  • REST APIs
  • MySQL
  • MongoDB
  • Pandas
  • NumPy
  • Matplotlib
  • Tableau

Tools

  • Git
  • Docker
  • Streamlit
  • Pytest
  • Jupyter
  • Kaggle
  • CUDA

03 · JOURNEY

Experience & education

Where I've been learning.

SMART INDIA HACKATHON

Team lead · InternAI

Led a team of 6 to build an internship recommender over 6,500+ listings, with Gemini-powered skill-gap analysis, resume feedback and learning roadmaps.

2023 – 2027

B.Tech, Computer Science and Engineering (AI) · Bennett University

Specialisation in Artificial Intelligence. CGPA 7.99 / 10. Coursework: data structures, algorithms, operating systems, DBMS, machine learning, generative AI.

CERTIFICATIONS

  • Deep Learning SpecializationDeepLearning.AI
  • Accelerated Computing in CUDA C++NVIDIA
  • Algorithmic ToolboxUC San Diego
  • Object-Oriented Data Structures in C++University of Illinois

04 · CONTACT

Let's build something.

I'm looking for AI/ML and Generative AI engineering roles. Email is the quickest way to reach me.

Download resume GitHub LinkedIn