Divyansh
Bansal

AI, agentic systems and whatever seems interesting enough to break.

The process

Ask the right question, stay with it, iterate until it holds, then make the useful thing real.

What I build

From a transformer written by hand to a voice agent that haggles with airlines, and the platform I build for AT&T.

• IntelliCode · Hybrid RAG

  Intelligence. Hybrid retrieval and AST analysis over a codebase. 121 tests · 92% coverage. Finding the right code means matching exact identifiers and intent at the same time, and without measurement there is no way to tell which retrieval stage is actually helping. A hybrid RAG pipeline: dense FAISS and sparse BM25, fused with reciprocal rank fusion and reranked by a cross-encoder, answering with Qwen2.5-3B on Hugging Face ZeroGPU. Beside it, an async-aware AST analyser covering 12 anti-pattern classes, and a security scanner for injection, weak crypto and hardcoded secrets. A labelled evaluation harness (MRR@5, Recall@k, NDCG) measures each stage's lift, and 121 tests at 92% coverage gate every change in GitHub Actions CI. Built with Python, Qwen2.5-3B, FAISS, BM25, Cross-Encoder Reranking, Hugging Face Transformers, GitHub Actions.

  IntelliCode · Hybrid RAG: CodeIntelliCode · Hybrid RAG: Live demo

• GPT From Scratch

  Models. A decoder-only transformer with no framework abstractions. 10.7M parameters · PyTorch. Using transformers every day is not the same as understanding one. The only way to be sure was to build it with nothing hidden. A decoder-only transformer of about 10.7M parameters in raw PyTorch: multi-head self-attention, positional encoding, layer normalisation and autoregressive generation, all written by hand. Training convergence and generation quality validated end to end on text corpora. Built with PyTorch, Python, Tokenisation, Sequence Modelling, Text Generation.

  GPT From Scratch: Code

• Loan Risk Intelligence

  Engineering. A stacking ensemble over ~1.8M LendingClub loans, SHAP-audited and drift-monitored. 0.9184 test AUC. Credit models trained on years of loans can quietly learn from the future. Across ~1.8M LendingClub loans, a 2016 underwriting regime shift leaks through any split that ignores time. A two-stage stacking ensemble (XGBoost and PyTorch) on year-stratified splits, audited with SHAP, and served from GCP Cloud Run (FastAPI, Docker) with live TreeSHAP explanations, a Streamlit dashboard and PSI/KS drift monitoring. A test AUC of 0.9184, validated with McNemar's test (χ² = 194.8, p < 0.0001). The API and the dashboard are live. Built with XGBoost, PyTorch, SHAP Explainability, MLflow, DagsHub, Drift Monitoring (PSI/KS), GCP Cloud Run, Docker, FastAPI, Streamlit.

  Loan Risk Intelligence: CodeLoan Risk Intelligence: APILoan Risk Intelligence: Dashboard

• DTU ERP Platform

  Product. An ML-powered ERP for higher education, live across the university. 1,200+ users. Preparing the university for accreditation took 25 days of manual work. An ML-powered ERP with 8+ modules for accreditation automation, analytics and networking, and a semantic-retrieval pipeline (vector embeddings, ChromaDB). I co-founded it as its product architect: I led the architecture and the frontend, trained the pilot department's faculty and staff, supported the rollout, and pitched it to the university administration. Preparation went from 25 days to 7 (about 72% less). 1,200+ users across the university, a paper in Wiley's Software: Practice and Experience, and the university now funds it. Built with ChromaDB, Vector Embeddings, Semantic Search, Docker, PostgreSQL.

  DTU ERP Platform: Paper · Wiley

• Agentic Reconciliation Platform

  Featured · Amdocs · AT&T. Agents for the language, a deterministic engine for every verdict. Built alone, for the AT&T account. Sole engineer · CI eval gate. Reconciliation on the AT&T account means proving that systems which should agree really do, table by table and row by row. That verdict has to be exact, so a language model can't be the one giving it. As the sole engineer: GPT-4.1 agents handle intake, NL-to-SQL and reporting, and a deterministic engine owns every verdict. LLM intake lifted completion from 71% to 96% of 126 test requests, at $0.004 each. A reproducible evaluation platform (12+ suites and a record/replay LLM proxy) that gates every merge request in CI at $0 API cost. It surfaced 38 ranked issues, the next two among them. Tables over 10k rows were being silently sampled. I rebuilt the read path with batched streaming, which made reconciliation exact at 1M rows per side on a synthetic benchmark: missed breaks went from 97,212 to 0. An agent-runtime defect failed every run. With it fixed, the NL-to-SQL agent reached 87.8% execution accuracy on 498 benchmark questions (a fixed-step pipeline: 67%), at $0.006 and 7.8 s p50 each. A self-healing loop that proves each format-noise fix by re-comparison before a human approves it. It cleared all noise on 16 of 18 test pairs while masking 0 real breaks across 17,842 out-of-sample cells. Every one of those numbers comes from a run pinned to its commit and its data, and it can be replayed. Along the way: technical discussions, solution demos and POCs with 40+ Amdocs and AT&T stakeholders, and their feedback turned into product iterations. Built with Python, FastAPI, Azure OpenAI (GPT-4.1), Agent Orchestration, NL-to-SQL, Structured Output Generation, Human-in-the-loop, CI/CD Eval Gates, GitLab CI, PostgreSQL, Redis.

  Also at Amdocs, Databricks migration · leading: Moving the subscriber-event pipeline from Oracle PL/SQL to Azure Databricks (about 80% of the way there): parsing inbound events, applying business rules and populating gold-layer tables, rebuilt in PySpark on a bronze/silver/gold medallion architecture with Unity Catalog. Along the way, a production stored-procedure run over 75M records a day went from 45 minutes to 12–15. Built with Databricks, PySpark, SQL, Oracle, Unity Catalog, Medallion Architecture.

• VocalisAI · Voice Agent

  Voice. A real-time voice agent that calls airlines for refunds. 91% success · 0 leaks. Airline refunds are won on the phone: long holds, scripted pushback, and knowing which passenger-rights rule applies where. An agent making that call also carries the passenger's personal details, which must never be said to the wrong party. A real-time voice agent (Pipecat, LangGraph, Deepgram) that negotiates refunds under Indian, UK/EU and UAE passenger-rights rules. Secrets never enter the model's context: placeholders are resolved only at speaking time, behind a deterministic streaming output guard. Jurisdiction-scoped hybrid retrieval (BM25, dense embeddings, cross-encoder reranking) for citing the regulations, lifting top-1 accuracy from 47% to 65% on 78 gold questions. An LLM judge that rated 29 of 30 calls 5/5, recalibrated with an anchored rubric: its agreement with an independent LLM labeller went from κ −0.05 to 0.73 on 30 held-out calls. A phonetic read-back fix for the phone line, which lifted task success over audio from 56% to 68% (25 calls each). 91% task success over 75 adversarial simulated calls in the text harness, with 0 sensitive-data leaks across them, and 1.86 s p50 (2.70 s p95) voice-to-voice latency on a simulated 8 kHz phone line. Built with Python, Pipecat, LangGraph, Deepgram, BM25, Cross-Encoder Reranking, Vector Embeddings, LLM-as-judge, Cohen's κ, Guardrails.

  VocalisAI · Voice Agent: Live demoVocalisAI · Voice Agent: Code

Stack

• AI & LLM Engineering: Azure OpenAI (GPT-4.1), LangChain, LangGraph, MCP, LiteLLM, Agent Orchestration, Guardrails, Human-in-the-loop
• Machine Learning & Deep Learning: PyTorch, TensorFlow, scikit-learn, XGBoost, LoRA / PEFT Fine-tuning, SHAP Explainability, Statistical Hypothesis Testing
• Retrieval & Search: FAISS, ChromaDB, RAG, Semantic Search, Vector Embeddings, BM25, Cross-Encoder Reranking
• Data Engineering: Databricks, PySpark, SQL, Oracle, Unity Catalog, Medallion Architecture, Apache Airflow. Certified Data Engineer Associate, Certified Generative AI Engineer Associate
• Data & Infrastructure: Python, FastAPI, Celery, Docker, GCP Cloud Run, PostgreSQL, Redis
• Voice & Language: Pipecat, Deepgram, Hugging Face Transformers, NL-to-SQL, Structured Output Generation, Text Generation, Sequence Modelling, Tokenisation
• Evaluation & MLOps: MLflow, DagsHub, CI/CD Eval Gates, LLM-as-judge, Cohen's κ, Drift Monitoring (PSI/KS), Model Versioning
• Tools & Languages: Git, GitHub Actions, GitLab CI, Claude Code, Pandas, NumPy, Streamlit, React

The early work.

• 2021, The beginning: B.Tech in Electronics and Communication Engineering at Delhi Technological University. CGPA 8.06.
• 2022 → 2023, First builds: Started turning ideas into systems.
• 2024, The ERP: 25 days → 7. We thought the process could be better. So we changed it.
• 2025, DTU ERP: 1,200+ users. A research paper. One university problem that got considerably larger.
• 2025, Amdocs · AT&T: AI software engineer on the AT&T account. Building its reconciliation platform alone: agents for the language, a deterministic engine for every verdict.

Still building.

Agentic AI on the AT&T account, and leading a migration from Oracle to Azure Databricks.