AI & Data Science Engineer · Deloitte US-India · ex-Cognizant · Gurugram, India
I build production AI systems for healthcare and complex documents, with a focus on extraction, agentic workflows, and evaluation.
Projects · Experience · Credentials · Activity · Contact
Independent projects from my resume. Each case study explains what I built, the evidence, and its limits.
| Project | What I built | Explore |
|---|---|---|
| DocLayout | Document conversion with layout-guided transcription, batch processing, and source-linked review. | Case study |
| Natural Language to SQL Agent | A LangGraph workflow with clarification, SQL validation, approval, and read-only SQLite/PostgreSQL execution. | Case study |
| LoRA QLoRA Fine-tune Studio | Dataset preparation, local adapter training, job monitoring, and evaluation against base models. | Case study |
| Hinglish Turn Detection | A 7.9M-parameter speech turn detector with a three-seed study and documented dataset limits. | Case study · Demo |
Seven case studies · decisions, evidence, and limitations
These seven studies cover independent tools and research. The Document AI repositories compare different approaches and remain separate applications.
DocLayout converts PDFs, scans, images, and supported office documents into structured Markdown and other outputs. PP-DocLayoutV3 ONNX analysis guides full-page GPT-6 Sol transcription. Accepted matches supply contours or fallback rectangles and relative reading order; unmatched content retains Sol geometry, and layout failures continue with an explicit Sol fallback.
The browser workbench supports single-file page selection and batches with three active document jobs. Users can compare rendered and raw output, ask document questions with GPT-6 Luna, and download Markdown, HTML, JSON, chunks, annotated page images, or annotated PDFs. Conversion also records usage metadata and API cost estimates.
After conversion, the GUI extracts authorization fields from completed raw Markdown with Sol at medium reasoning effort. It saves schema-validated JSON and SQLite records for persistent review, with source quotes linked to PDF blocks and unresolved evidence flagged for review. Classification is off by default. Field extraction belongs to the GUI workflow; the CLI, library, and HTTP API remain conversion interfaces.
I built DocLayout on the Apache-licensed Marker project and added the document workbench, model-backed extraction, document chat, exports, and usage cost estimates. DocLayout is an independent application and does not imply upstream endorsement.
View the DocLayout repository and installation guide
GroundMark parses PDFs and images page by page with GPT-6 Sol. Local PP-DocLayoutV3 ONNX analysis provides region hints before Sol reads the full page. Accepted matches supply geometry and relative reading order, using mask-derived contours where available and preserving Sol's text. Unmatched content keeps Sol geometry; unavailable layout analysis falls back to Sol with diagnostics.
Its Streamlit interface and CLI export Markdown, HTML, JSON, annotated PDFs, annotated page images, and Markdown ZIPs with figure crops. Successful pages remain available when other pages fail, with page diagnostics explaining partial results. GPT-6 Luna answers questions over successfully parsed pages with inline page references, local quote checks, and a separate verification call.
The parser retains page, block, geometry, and layout provenance for review. Detailed layout remains a separate, experimental Sol option, off by default. Its earlier transcription evaluation does not measure the V3 integration.
GroundMark architecture and workflow diagrams
View the GroundMark repository
Paperplane flow: Document → OCR → Extraction → Evaluation.
Scans, handwriting, tables, and native documents call for different processing strategies. Readable text alone is insufficient when extracted fields and answers need evidence a person can inspect.
The lab contains eight separate implementations for local OCR, multimodal parsing, schema extraction, annotated review artifacts, and GraphRAG question answering. They keep source evidence distinct from model interpretation, make local and cloud processing comparisons explicit, and treat document Q&A separately from parsing.
The repositories expose processing paths, output contracts, and review artifacts. They do not establish a shared accuracy benchmark or parity with a commercial extraction service, and privacy and provider boundaries differ by implementation.
| Implementation | Approach |
|---|---|
| Paperplane | Explicit parsing engines with grounded Markdown, JSON, and document organization. |
| Document Intelligence Agent | GraphRAG retrieval, verification, and cross-document comparison. |
| Grounded DocParse | Mixed-format parsing with immutable source spans and evidence-linked extraction. |
| OpenAI + RapidOCR | Local OCR and layout evidence followed by multimodal refinement. |
| OpenAI Agentic Document Extraction | Model-cascade extraction, segment verification, and disputed-field resolution. |
| NaviDC-OCR Studio | Local GPU OCR with optional cloud-based semantic field extraction. |
| LiteParse Agentic Extraction | Layout reconstruction with model OCR and targeted hard-region verification. |
| ADE workflow example | Bounded LangGraph parsing that produces Markdown, HTML, and annotated pages. |
Flow: Dataset → Prepare → Train → Compare.
Local fine-tuning depends on dataset compatibility, GPU capacity, run configuration, recovery, and adapter review. This Streamlit workflow covers dataset preparation, model inspection, parameter-efficient training, job monitoring, and base-versus-adapter comparison. It saves portable PEFT adapters and maintains a persistent local job queue.
Runs are validated before launch, GPU workers are isolated, and the queue processes one run at a time. The app supports cancellation, logs, and checkpoint recovery, alongside a CUDA-free synthetic showcase.
The repository includes the training implementation, example datasets, workflow documentation, and a read-only showcase. Adapter quality depends on the selected data, model, and recipe. It is a single-user local experiment environment and does not provide hosted or distributed training; the separate showcase performs no training.
View the LoRA Fine-tune Studio repository · Read the LoRA portfolio study
Flow: Prompt → Generate → Evaluate → Select.
A prompt that looks better on one example may perform worse elsewhere. This LangGraph workflow generates diverse prompt candidates and evaluates them against the same selected benchmark, keeping an elite pool. Streamlit shows scores, Pareto trade-offs, per-case results, and downloadable history so the comparisons can be reviewed.
The workflow supports weighted, Pareto, and hybrid selection. It caches exact prompt scores and retains the strongest candidate across generations.
Candidate generation and judge-based scoring can be inspected. A higher ju 10BC8 dge score is an experimental result; it does not establish improvement outside the selected benchmark. Conclusions depend on judge quality and benchmark coverage. Evaluation runs sequentially, with run state held in memory, and the repository has no automated test suite.
View the Prompt Optimizer repository · Read the Prompt Optimizer portfolio study
Flow: Video → Evidence → Retrieval → Answer.
Reprocessing a complete video for every task repeats work over the same speech and visual evidence. This local Streamlit workspace uses Adversal’s remote MCP video analysis and reuses returned Markdown, timestamps, and frames. Qdrant scopes retrieval to the selected video for questions and generated documents.
The app saves completed source artifacts, request IDs, and job state for monitoring and recovery. Generated answers include their source material.
The repository includes job lifecycle handling, evidence indexing, export workflows, and focused tests using service mocks. Adversal supplies the underlying video-understanding service, and model calls also use external services. Each browser session has one active video workspace; generated documents and chat history stay within that session.
View the Video Summarizer repository · Read the Video Summarizer portfolio study
A pause is not always a finished utterance, and false-complete decisions can cause a voice agent to interrupt the speaker. The study compares Whisper-tiny-based audio classifiers, pooling and augmentation strategies, an audio-text variant, and a three-seed finalist with validation-calibrated thresholds.
It tracks false-complete rate alongside recall and F1, selects the final architecture and threshold from validation results before held-out evaluation, and compares audio-only and audio-text approaches for both quality and live inference overhead.
The selected checkpoint achieved a 9.84% false-complete rate and 83.26% recall on 4,890 held-out examples. Recall missed the desired 85% target. The data lacks verified Hinglish/code-switch labels and speaker IDs, and much of the training audio is synthetic, so this is not a speaker-disjoint benchmark of real human Hinglish conversations.
View the Hinglish Turn Detection repository · Try the Hinglish Turn Detection demo · Read the Hinglish portfolio study
Browse the complete project catalog
| Project | Focus | What it does |
|---|---|---|
| DocLayout | Document AI · Extraction | Converts documents with PP-DocLayoutV3 and Sol; its browser workbench adds batch processing, authorization fields, and persistent source-linked review. |
| GroundMark | Document AI · Extraction | Combines PP-DocLayoutV3 geometry with Sol transcription, reviewable exports, partial-page diagnostics, and cited document chat. |
| LoRA Fine-tune Studio | Model Training · Fine-Tuning | Prepares datasets, trains local LLM adapters, monitors jobs, and compares adapters with base models. |
| Tool-Using Browser Agent | Agentic AI · Browser Automation | Plans and performs Playwright actions with memory and human approval before sensitive steps. |
| Self-Improving Prompt Optimizer | Evaluation · Prompt Optimization | Generates and compares prompt candidates with multi-objective LLM-as-judge scoring. |
| NL2SQL Agent | Agentic AI · Data Systems | Produces reviewed SQL for SQLite and PostgreSQL with audit trails and local or hosted models. |
| Autonomous Coding Agent Crew | Agentic AI · Developer Tools | Coordinates planning, specialists, review, testing, debugging, documentation, and quality gates. |
| Multi-Agent Debate Decision System | Multi-Agent AI · Decision Support | Runs moderated debates with structured judging, confidence, recommendations, and risks. |
| Multi-Agent Research Assistant | Multi-Agent AI · Research | Coordinates planning, parallel research, critique, reflection, and cited report writing. |
| Local-First Knowledge Base Agent | Local AI · Knowledge Systems | Indexes personal knowledge for cited answers, temporal reasoning, and local knowledge graphs. |
| Intelligent Personal Finance Agent | Local AI · Personal Finance | Ingests statements, learns category corrections, and keeps the ledger on the user’s machine. |
| Document Intelligence Agent | Document AI · GraphRAG | Uses LangGraph and ArcadeDB for extraction, community detection, retrieval, and verification. |
| Autonomous Job Application Agent | Agentic AI · Workflow Automation | Collects and scores jobs, drafts tailored content, and tracks applications locally with human review. |
| AutoTabML Studio | Machine Learning · AutoML | Turns tabular data into trained, evaluated, deployable models through reproducible UI and CLI workflows. |
| Codebase Understanding Agent | Agentic AI · Developer Tools | Clones, scans, summarizes, and explains repositories, then answers codebase questions through chat. |
More repositories
- computer-use
- Grounded DocParse — Mixed-format parsing with immutable source spans and evidence-linked extraction.
- Paperplane — Parse documents into grounded Markdown, JSON, and review artifacts.
- local-ai-chat-studio
These are team and system results from internal employer evaluations. My contribution is stated separately. Client identities, internal project names, source data, prompts, schemas, thresholds, scoring details, and proprietary code are omitted. Related public projects demonstrate engineering patterns; they do not reproduce these measurements.
- 80–81% to 92%+ structured-extraction accuracy. Deloitte; recurring 100-file internal evaluation runs. I designed grouped extraction, confidence-aware four-pass extraction, retries, validation, and routing.
- 38% to 80% browser task completion. Deloitte; the same 200-task internal evaluation. I contributed Milvus retrieval, reranking, and failure-aware routing.
- 79% to 88% warranty-classifier recall. Cognizant; reported internal classifier evaluation. I retrained and tuned Random Forest, compared XGBoost, and retained Random Forest based on results.
All six outcomes and evaluation contexts
| Outcome | Evaluation context | My contribution |
|---|---|---|
| 95%+ fax-classification accuracy | Deloitte; measured across 500-file bulk batches. | Worked on Azure Databricks classification and routing for fax intake. |
| 80–81% to 92%+ structured-extraction accuracy | Deloitte; recurring 100-file internal evaluation runs. | Designed grouped extraction, confidence-aware four-pass extraction, retries, validation, and routing. |
| 38% to 80% browser task completion | Deloitte; the same 200-task internal evaluation. | Contributed Milvus retrieval, reranking, and failure-aware routing. |
| ~40% lower browser-agent prompt-token use | Deloitte; internal comparison against raw DOM observations. | Built accessibility-tree snapshots and compressed observations. |
| 90% to 99% policy-entity extraction accuracy | Deloitte; the same internal benchmark. | Iterated prompts, implemented canonical comparison, and expanded evaluation. |
| 79% to 88% warranty-classifier recall | Cognizant; reported internal classifier evaluation. | Retrained and tuned Random Forest, compared XGBoost, and retained Random Forest based on results. |
Read my professional experience · Review the sanitized outcome boundaries
- Deloitte US-India · AI & Data Science Engineer · July 2025 – Present · Gurugram, India
- Cognizant Technology Solutions · Associate Data Scientist · September 2022 – May 2025 · Noida, India
- AiEnsured · Machine Learning Engineer Intern · July 2021 – August 2021 · Remote, India
Project contributions, engineering decisions, and architecture snapshots
July 2025 – Present · Gurugram, India
A production Azure pipeline processes incoming healthcare fax packets from classification through extraction, validation, routing, and recovery. Handwriting, nonstandard forms, business rules, and a 117-field schema made a single extraction call unreliable.
-
Used Azure Content Understanding to generate Markdown, then Azure OpenAI to extract structured values.
-
Designed the production revision around seven calls grouping related fields, replacing one request for all 117 fields.
-
Designed confidence-aware four-pass extraction, retries, validation, and explicit recovery outcomes.
-
Migrated GPT-4.1 prompts to GPT-5.2 and checked the revision with manual review and regression tests.
-
Kept classification and extraction separate, producing RPA-ready CSV/JSON and annotated PDFs. Grouping related fields across seven calls and running confidence-aware extraction passes are separate parts of the approach.
-
The grouped extraction revision went live in September 2026.
Read about prior-authorization document processing
Three separate projects covered document integrity, out-of-network claims, and referral patterns.
- Partnered with clinical, operational, and business stakeholders to map workflows and document PHI-aware solution designs, delivery risks, and implementation guidance.
- Built a seven-agent LangGraph and GPT-4o Vision workflow that combined visual, metadata, and semantic document checks with NPI/EIN validation and deterministic 0–100 risk scoring.
- Developed out-of-network claims-analysis dashboards with deterministic risk signals, evidence review, and investigator-facing reporting.
- Developed referral-pattern analysis with agentic orchestration, approval gates, evidence review, and report generation.
Read about healthcare integrity and fraud analytics
- Contributed Milvus retrieval, reranking, and failure-aware routing to a computer-use workflow.
- Built a Playwright MCP tool using accessibility-tree snapshots and compressed-vision context instead of raw DOM observations.
- Task completion and prompt-token use were evaluated as separate outcomes.
Read about computer-use and multi-agent reasoning
Iterated prompts, implemented canonical comparison, and expanded evaluation for structured policy-entity extraction used in care-management decision support.
Read about policy-entity extraction
September 2022 – May 2025 · Noida, India
- Improved an automotive warranty workflow handling roughly 1,800–2,200 claims per weekday and approximately 1,000 per weekend day.
- Retrained and tuned its existing Random Forest classifier, compared XGBoost, and retained Random Forest because XGBoost did not outperform it.
- Prioritized recall because missing a legitimate claim carried a higher business cost than additional review.
- Migrated legacy LUIS intent and entity processing to Azure Conversational Language Understanding.
- Helped move expensive downstream processing out of the FastAPI request path through Azure Blob Storage and Azure Functions after HTTP 504 timeouts under a 30-second SLA.
- Built Power BI analytics, model-behavior, and drift dashboards.
- CLU calls remained in FastAPI; separate downstream processing moved through Azure Blob Storage and Azure Functions.
Read about warranty processing
For a separate FMCG engagement, I built a conversational reordering workflow using AWS Lex, Azure OpenAI, AWS Lambda, and Amazon S3.
July 2021 – August 2021 · Remote, India
- Supported CNN-based object-detection work, including code optimization and error analysis.
- Implemented regression and classification models and contributed to feature-engineering experiments.
- Prior authorization: Fax documents → type and urgency classification → eligible documents → Azure Content Understanding Markdown → grouped Azure OpenAI extraction → validation and business rules → structured output or review.
- Computer use: Retrieved knowledge and browser observations → multi-agent reasoning → tool execution → explicit routing and failure handling.
- Warranty processing: Standard claim scoring remains separate from NLP processing; accepted payloads move through Blob Storage to Azure Functions for downstream work.
LLM architectures and agentic workflows · Retrieval and knowledge systems · Multimodal document intelligence · Evaluation and production ML · Event-driven cloud and backend systems.
How I use these capabilities: professional work, personal projects, and learning
- LLM architectures and agentic workflows: context engineering and prompt hardening for structured workflows; separate agentic systems add tools, planning, approval gates, memory, retries, and explicit failure handling.
- Retrieval and knowledge systems: hybrid retrieval, reranking, graph traversal, citations, and local-first memory for grounded answers.
- Multimodal document intelligence: layout-aware extraction from scanned and handwritten faxes using Azure Content Understanding and Azure OpenAI, with grouped fields, confidence-aware retries, schema validation, and human review.
- Evaluation and production ML: Random Forest retraining, XGBoost comparison, recall-led evaluation, drift dashboards, prompt regression checks, fixed internal benchmarks, and personal experiments in LLM-as-judge evaluation and LoRA/QLoRA.
- Event-driven cloud and backend systems: Python and FastAPI services with downstream Azure Blob Storage and Functions processing; schema-first contracts, local/cloud model routing, reproducible environments, and deployment controls in public projects.
These stages appear across separate projects. Each system uses the stages its problem needs.
Data → Retrieval → Model → Agent → State → API → Deployment → Evaluation → Observability → Iteration
The measured outcomes and linked project stories show how these capabilities were applied and evaluated.
| Tools | Applied context |
|---|---|
| Azure OpenAI, Azure Content Understanding, Azure Databricks | Grouped LLM extraction over CU Markdown with confidence-aware four-pass extraction and validation. |
| LangGraph, GPT-4o Vision, Streamlit | Specialist document checks, deterministic risk scoring, evidence review, and investigator-facing reporting. |
| Milvus, Playwright, MCP | Retrieval, reranking, and efficient browser observations for multi-agent reasoning. |
| Python, FastAPI, Azure App Service, Blob Storage, Functions, CLU, Random Forest, XGBoost, Power BI | Warranty classification, NLP migration, downstream processing, analytics, and drift monitoring. |
| AWS Lex, Azure OpenAI, Lambda, S3 | Conversational B2B reordering for a separate FMCG engagement. |
| Tools | Public implementation |
|---|---|
| LangChain, Anthropic, Gemini, Ollama, Pydantic, SQLite, PostgreSQL | NL2SQL Agent |
| PyTorch, Transformers, PEFT, TRL | LoRA Fine-tune Studio |
| Chroma | Google OKF implementation |
| ArcadeDB, LangGraph | Document Intelligence Agent |
| Docker, GitHub Actions, Qdrant | Video Summarizer |
I am learning about the following areas and do not claim production expertise in them.
- Agentic AI and orchestration: LangGraph + MCP, AutoGen, CrewAI, multi-agent systems
- Reasoning models and post-training: test-time compute scaling, GRPO, DPO
- Evaluation and AI governance: LLM-as-a-Judge, DeepEval, NeMo Guardrails, red-teaming, precision-recall curves
- Edge AI and high-throughput inference: vLLM, SGLang, Ollama, FP4/FP8 quantization, AWQ
- Agentic RAG and GraphRAG: Neo4j, LlamaIndex Workflows, vector databases, hybrid search
- Multimodal and vision-language AI: VLMs, Document AI, audio-to-audio systems
- Cloud events and observability: Azure Event Grid, Application Insights
My background and how I work
At Deloitte, I work on healthcare document processing, retrieval and computer-use workflows, and healthcare integrity analytics. Previously, at Cognizant, I improved production machine-learning and NLP systems, built conversational AI, and helped move expensive processing into event-driven Azure services.
My work began in ML, NLP, and data science, expanded into conversational AI, and now includes GenAI and agentic AI. Evaluation and operational reliability have stayed central throughout.
I start by researching existing solutions and comparing model quality and cost. I inspect early results manually and regression-test changes, keeping source evidence available for review. After delivery, I monitor the systems. I prefer typed boundaries, explicit failure states, and human review where needed, and I keep my claims within what the evidence supports.
Forward-deployed AI engineering · ongoing learning
I am building on professional experience and independent projects to learn forward-deployed AI engineering. The path is ongoing and does not represent a completed curriculum, an employment title, or ownership of the full customer lifecycle.
- Problem discovery (professional experience): Partnered with clinical, operational, and business stakeholders to map workflows and document PHI-aware solution designs.
- Solution definition (professional experience): Documented delivery risks, constraints, and implementation guidance connecting workflow requirements to engineering decisions.
- AI prototyping (professional experience): Iterated structured extraction and prompts, compared model behavior, and inspected intermediate Markdown for omissions and hallucinations.
- Retrieval systems (professional experience): Contributed Milvus retrieval, reranking, and failure-aware routing to computer-use and multi-agent reasoning.
- Agentic systems (professional experience): Built LangGraph specialist checks combining visual, metadata, and semantic evidence with review and reporting.
- State and recovery (personal projects): Built persistent training queues and checkpoint recovery, plus resumable video-analysis requests and saved job state.
- Deployment and integration (professional experience): Helped decouple expensive warranty processing through Azure Blob Storage and Azure Functions.
- Operational monitoring (professional experience): Built analytics and model-drift dashboards for changing claim behavior and service issues.
- Evaluation and iteration (professional experience): Used manual review and regression testing with confidence-aware extraction, retries, and validation.
- M.Tech in Data Analytics and Decision Sciences, Indian Institute of Information Technology Kurnool, October 2020 to June 2022.
- B.Tech in Computer Science Engineering, Maulana Azad National Urdu University, August 2015 to June 2019.
Degree coursework
- M.Tech: machine learning, deep learning, NLP, computer vision, and statistics.
- B.Tech: data structures and algorithms, engineering mathematics, and web development.
Certified Partner Specialist Gemini Enterprise Deployment
Issued October 7, 2026
Verify on Credly
Claude Certified Architect - Professional
Issued October 14, 2026
Verify on Credly
Claude Certified Developer - Foundations
Issued October 14, 2026
Verify on Credly
Claude Certified Architect - Foundations
Issued September 30, 2026
Verify on Credly
Claude Certified Associate - Foundations
Issued August 31, 2026
Verify on Credly
Skill badges and course certificates · 17 credentials
- Accelerate Development with Antigravity: completed October 7, 2026
- Add Agents to Gemini Enterprise: completed October 7, 2026
- Deploy Gemini Enterprise with Workspace Data Sources and Model Armor: completed October 7, 2026
- Govern Agent Access with Gemini Enterprise Agent Platform: completed October 7, 2026
- Claude Code 101: completed August 18, 2026
- Building with the Claude API: completed July 6, 2026
- Claude 101: completed March 10, 2026
- AI Fluency: Framework & Foundations: completed March 11, 2026
- Machine Learning Specialization: DeepLearning.AI / Stanford via Coursera
- Advanced Learning Algorithms: DeepLearning.AI / Stanford via Coursera
- Supervised Machine Learning: DeepLearning.AI / Stanford via Coursera
- Unsupervised Learning & Recommenders: DeepLearning.AI / Stanford via Coursera
- Deep Learning A-Z™: Udemy
- Deep Learning A-Z™ Hands-On: Udemy
- Machine Learning A-Z™: Udemy
- Data Science for Professionals: Udemy
- SQL for Data Science: Coursera
View the complete education and credential portfolio
View contribution activity on my portfolio.
Statistics, languages, streaks, and the advanced dashboard
Open the live GitHub dashboard for repository discovery, releases, collaboration, contribution, and traffic snapshots.
Advanced GitHub dashboard
Contribution history
I’m based in Gurugram, India (IST, UTC+05:30), and open to remote-first production AI and GenAI engineering roles. I’m also available for focused consulting projects in multimodal document intelligence, LLM extraction, RAG, agentic workflows, and evaluation.
Production AI · Document Intelligence · Agentic Systems · Evaluation
Repository documentation and governance
The main branch is the source of record. See the changelog.
- Profile and outcome boundaries:
README.md,docs/sanitized-outcomes.md - Profile maintenance: edit this README directly. The existing GitHub Actions workflows refresh statistics, dashboard exports, contribution history, and arcade animations.
- Governance:
CONTRIBUTING.md,SECURITY.md,CODE_OF_CONDUCT.md
Made with ❤️ by Ahmad Mujtaba