Open to opportunities

Harshal
Bhambhani

AI engineer at OTO Capital, building production voice agents. BITS Pilani Hyderabad graduate (EEE, 2026). From NLP pipelines at Standard Chartered to live voice agents, multi-agent systems, knowledge-graph RAG, and real estate analytics — I build things that get deployed and actually used.

90%
Audit time reduced
19
Projects built
13+
Language models integrated
3
AI agent systems built
01 — About

Engineer by training,
builder by instinct

I started in Electrical and Electronics Engineering at BITS Hyderabad, but what genuinely pulled my interest was the Computing and Intelligence minor I picked up alongside — machine learning, AI, database systems, statistics. That curiosity has driven everything since.

My first real-world exposure was a machine learning engineering internship at DG Liger Consulting, where I built my first RAG chatbot and learned the hard gap between theory and production. Then came a Data Science internship at Standard Chartered GBS — six months of building a production NLP pipeline that real auditors depended on every day.

Since August 2026 I have been an AI Engineer at OTO Capital, building live voice agents for lending clients including Jana Small Finance Bank, Swarnodhayam MFI, Fusion MFI, Seed Fincap and Satin, after a short consulting engagement at Goel Exports on their Odoo workflow.

I don't just build models. I build systems that get deployed, used, and measured. The feedback loop between shipping something and watching it change how people work is what drives me forward.

Outside the code: double-sport state representative (basketball and cricket), volleyball gold medalist, and former captain of logistics teams managing 20+ people at large-scale campus events.

Python Gemini Live LangGraph Playwright Pipecat Odoo LangChain FastAPI DistilBERT SBERT FAISS NLLB-200 Whisper Qwen 7B SQL Power BI Docker (coursework) AWS (coursework) Pandas Scikit-Learn PaddleOCR Dataiku DSS Streamlit XGBoost LightGBM SHAP Neo4j spaCy RoBERTa Groq LLaMA 3 Mistral Gensim dbt PostgreSQL Flutter Firebase Selenium BeautifulSoup Flask InLegalBERT Vespa OpenAI API Vertex AI LLM-as-Judge A/B Testing Context Caching Exotel asyncpg GCP AWS RDS Ollama ChromaDB
BITS Pilani, Hyderabad Campus
B.E. Electrical & Electronics Engineering
Minor: Computing & Intelligence (ML, AI, DB Systems)
2022 — 2026 CGPA 6.9
MDS Public School
Class XII — CBSE
2021 92.2%
St. Paul's Convent School
Class X — CBSE
2019 91.2%
Coursework Highlights
Machine Learning · Statistics & Probability for Analytics · Database Systems · Artificial Intelligence · DSA · Operating Systems · OOP · Computer Programming
02 — Experience

Where I've built things

Aug 2026
Present
AI Engineer
OTO Capital · Bangalore
  • Built a production Gemini Live voice agent on Pipecat and FastAPI for collections calls, delivered to Jana Small Finance Bank, Swarnodhayam MFI, Fusion MFI, Seed Fincap and Satin in 10 languages
  • Moved negotiation policy out of the prompt into a deterministic Python gate that approves or refuses the model's tool calls before it speaks
  • Cut per-turn prompt cost 78% (16,558 to 3,596 tokens) by moving situational context into tool responses
  • Pinned each of 49 live-call defects with a regression check (201 at the first milestone, 845 gate checks by October), backed by a seeded simulator that runs 500+ borrower calls per change
  • Measure agent variants with A/B arms assigned by SHA-256, with 33 telemetry fields per call joined to payments
  • Moved post-call analytics from Gemini to OpenAI gpt-6-luna with a per-call Gemini fallback: live for Satin at about ₹0.22 vs ₹0.58 per call across 8,402 calls, with the queue step matching Gemini on 99.1% of 3,110 past calls
  • Mined 36,091 call rows down to 12 curated scenes (same-day promises converted at 14.1% vs 7.3% for next day); audited 26 human collector calls against RBI norms, flagging 17 practices and locking 9 as prohibited
Jul 2026
Aug 2026
Data & Analytics Consultant
Goel Exports
  • Mapped an exporter's manual order-to-cash process into a documented 4-phase Odoo 19 workflow
  • Configured field-level data rules across 6 process stages, agreeing each with the operations owners
Jul 2025
Dec 2025
Data Science Intern
Standard Chartered Global Business Services · Bangalore
  • Engineered NLP pipelines on Dataiku DSS to process 100+ page LMA agreements, cutting manual audit times by 90% for compliance teams
  • Fine-tuned DistilBERT with WeightedTrainer (20x class weight boost) achieving 98.6% recall on rare legal clauses despite 90%+ irrelevant text
  • Fine-tuned SBERT with MultipleNegativesRankingLoss for semantic search — retrieved gold-standard compliance language via cosine similarity
  • Deployed confidence thresholding (minimum correct confidence method) filtering 95%+ false positives for audit-grade precision
  • Built AutoTransFlow: multilingual PDF translation using doc-layout-yolo + NLLB-200, preserving original document layout without external APIs
Jun 2024
Jul 2024
Machine Learning Engineer Intern
DG Liger Consulting · Gurgaon
  • Built production RAG chatbot: LangChain loaders → RecursiveCharacterTextSplitter → all-MiniLM-L6-v2 embeddings → FAISS → StableLM Zephyr 3B
  • Implemented ConversationalRetrievalChain with ConversationBufferMemory for multi-turn dialogue with proprietary PDF documents
  • Integrated speech and NLP models for analytics automation; scraped and processed 500+ articles via BeautifulSoup
03 — Projects

Things I've shipped

📜
↗
RegRAG — RBI Banking-Regulation RAG Benchmark
Question answering over RBI Master Directions with page citations, and a benchmark of six RAG designs on the same questions. An evidence gate abstains with "not found" when the documents do not support an answer. Early development.
RAG LLM Evals Guardrails Ollama
Hybrid retrieval recall@5 0.92 on 26 questions · 6 pipelines compared · full benchmark in progress
🎙️
↗
Collections Voice Agent — Deterministic Negotiation Gate
A production Gemini Live agent for loan-collection calls in 10 languages. Negotiation policy lives in a Python gate that approves or refuses each tool call before the model speaks. Delivered for Jana Small Finance Bank, Swarnodhayam MFI, Fusion MFI, Seed Fincap and Satin.
Gemini Live Pipecat LLM Evals A/B Testing
Prompt tokens per turn down 78% (16,558 to 3,596) · 49 live-call defects pinned by 201 regression checks
🎙️
↗
Voice Agent Audit — Evaluating a Hinglish Voice Agent
Evaluation kit for a Hinglish payment-reminder voice agent: rule checks, an LLM judge checked against human labels, and hybrid retrieval over a pgvector policy corpus, served with FastAPI, Redis and Langfuse tracing.
LLM Evals pgvector FastAPI Langfuse
43.5% fewer prompt tokens · 119/120 correct actions · retrieval recall@3 0.53
🛠️
↗
Sanwaad — Multi-Agent Complaint Resolution
Multi-agent system that turns public complaints into resolved cases. LangGraph pipeline with hybrid RAG, policy-grounded replies, and human approval on every money action. 472 tests and trace-level evals.
LangGraph Gemini Hybrid RAG FastAPI
15/15 trajectory scenarios passed · 0 safety violations · retrieval recall@5 1.00
🏦
↗
LMA Clause Identification Tool
Production NLP pipeline on Dataiku DSS identifying legal clauses in loan agreements. Fine-tuned DistilBERT + SBERT semantic search with confidence thresholding.
DistilBERT SBERT Dataiku DSS NLP
↑ 90% reduction in audit time · 98.6% recall
🌐
↗
AutoTransFlow
Multilingual document translation preserving original PDF layout. doc-layout-yolo detects text blocks with bounding boxes; NLLB-200 translates 200 languages locally.
NLLB-200 YOLO Computer Vision PDF AI
Layout-preserving translation · 200 languages · Zero external API calls
🤖
↗
RAG Chatbot
Production retrieval-augmented generation chatbot. LangChain pipeline with FAISS vector database, all-MiniLM-L6-v2 embeddings, and StableLM Zephyr 3B for grounded answers.
LangChain FAISS StableLM RAG
Multi-turn dialogue · Grounded answers · PDF knowledge base
🏥
↗
Medical Bill OCR & Fraud Detection
End-to-end datathon pipeline: OpenCV preprocessing → PaddleOCR → Qwen 7B extraction → 4 automated fraud checks (IQR, reconciliation, SHA-256 hashing, pattern analysis).
PaddleOCR Qwen 7B Fraud Detection OpenCV
90%+ extraction accuracy · 4 fraud detection checks
🎙️
↗
Multilingual ASR System
Zero-shot Urdu speech recognition using Whisper. Audio preprocessing pipeline with pydub and librosa, IndicBERT MLM corrector for post-ASR error fixing.
Whisper IndicBERT pydub Speech AI
↓ 14% Word Error Rate · Zero-shot Urdu recognition
📈
↗
Market Analysis Using LLMs
Financial news sentiment pipeline. BeautifulSoup scraping → two-pass BART summarisation → RoBERTa sentiment → SMOTE balancing → Flask API with Streamlit dashboard.
BART RoBERTa Flask SMOTE
500+ articles processed · Real-time sentiment API
🧭
↗
OpsAgent — Autonomous Browser Operator
AI agent that completes back-office tasks in a real browser using a plan-act-verify loop. Approval gates enforced in code, and an independent verifier that re-reads the ERP without an LLM.
Gemini Playwright FastAPI AI Agent
11/12 eval scenarios passed · about $0.03 per run · approval gates in code
📊
↗
Audit Insight — NLP Analytics
End-to-end NLP pipeline for audit text analysis. RoBERTa sentiment classification, Gensim LDA topic modeling, KMeans clustering with t-SNE visualization, and Streamlit dashboard.
RoBERTa spaCy LDA Streamlit
Sentiment + topic + clustering · Interactive dashboard
⚡
↗
Career Copilot 2.0
AI-powered job search operating system. RAG matching with FAISS, response prediction, hidden market radar, ATS optimization, and 11 intelligent modules. Built with Streamlit + LangChain + Groq LLaMA 3.
LangChain FAISS Groq LLaMA 3
11 AI modules · Complete career operating system
✈️
↗
AirFare-ML
End-to-end airline ticket price prediction. 6 regression algorithms benchmarked with Bayesian hyperparameter optimization. LightGBM achieves 93.9% R² with ₹689 MAE.
LightGBM XGBoost scikit-learn SHAP
93.9% R² · ₹689 MAE · 6 models benchmarked
🎨
↗
Hastkala Setu
Dual-component BITS design project: ML pipeline for cross-cultural conflict resolution (MANOVA, XGBoost, SHAP) + Flutter mobile app for Indian handicraft artisans with Hindi voice input.
XGBoost SHAP Flutter Firebase
0.938 AUC · 776 subjects · Field-tested app
🚦
↗
Road Safety NLP Analysis
NLP-based classification of road accident severity from news articles. Multi-source scraping pipeline with spaCy NER for location extraction and severity classification across 8 Hyderabad news outlets.
spaCy BeautifulSoup NLP NER
8 news sources · Severity classification · NER location extraction
🏗️
↗
HydRERA Analytics
End-to-end real estate analytics platform. ELT pipeline (Python → PostgreSQL → dbt) producing builder reliability scorecards, supply-demand tracking, price fairness analysis, and delay prediction with SHAP.
dbt PostgreSQL Power BI XGBoost
4 analytical modules · Builder scorecards · Delay prediction
🔗
↗
Knowledge Graph RAG for Legal AI
Hybrid KG+RAG system for Indian Supreme Court case analysis. Neo4j legal knowledge graph with 1,000 cases, Vespa hybrid search (BM25 + InLegalBERT), and IRAC-constrained LLM reasoning with Mistral 7B.
Neo4j InLegalBERT Vespa Mistral
1,000 cases · IRAC reasoning · KG vs GraphMERT benchmark
🧬
↗
StemCells DBMS
Full-stack donor, patient, hospital, and staff management system. Flask + MySQL backend with dynamic dashboard that tracks entity counts in real-time as new registrations occur.
Flask MySQL HTML DBMS
Dynamic dashboard · CRUD operations · Real-time counts
04 — Contact

Let's connect

I'm open to opportunities in AI engineering, voice agents, data science, and GenAI development. If you're building something interesting or want to discuss the work — I'd love to talk.

Quick summary
Role
AI Engineer
Location
Bangalore
Status
Open to opportunities
Focus
Voice agents · NLP · GenAI