Tarun Kumar Reddy Nallagari
Professional Summary
AI/ML Engineer with 4+ years of experience building retrieval-augmented generation systems, agentic AI workflows, and probabilistic forecasting pipelines across financial services and consumer goods manufacturing. Designed and deployed a multi-stage document-intelligence platform at Ally Financial handling financial and compliance PDFs using hybrid BM25 and HNSW retrieval, cross-encoder reranking, and LLM-as-a-Judge evaluation with Claude Haiku, Sonnet, and Opus. At AB InBev India, engineered a neural meta-ensemble combining ARIMA, Prophet, XGBoost, LightGBM, LSTM, and TFT for brewery empties-return forecasting, backed by Monte Carlo simulation and Value-at-Risk analysis to support capacity and cost-risk decisions. Proficient in Python, PySpark, FastAPI, Snowflake, PostgreSQL, AWS SageMaker, and Docker. Holds an M.S. in Data Science from the University at Buffalo and an AWS Certified Cloud Practitioner credential.
Technical Skills
Work Experience
- Architected a secure document-intelligence workflow for financial and compliance PDFs, connecting Next.js, FastAPI, role-based access control, and Amazon S3 to support batch processing of 1 to 1,000 documents per request.
- Built asynchronous ingestion pipelines using SQS and Redis queues with parallel workers to extract text, tables, and document structure, then generate hierarchical chunks and embeddings for PostgreSQL/pgvector indexing.
- Developed hybrid retrieval combining BM25 keyword search and HNSW vector search with reciprocal-rank fusion, cross-encoder reranking, metadata filters, and parent-context expansion to surface the most relevant evidence.
- Implemented fast-path routing and multi-step agent orchestration using LangGraph to handle document extraction, summarization, comparison, calculation, and risk-analysis request types within a single agentic loop.
- Designed citation validation and LLM-as-a-Judge groundedness checks against golden-dataset regression reviews to flag unsupported answers and drive iterative retrieval and prompt refinement.
- Benchmarked cost-aware LLM routing across 1,000 pull-request-based software tasks, evaluating Claude Haiku, Sonnet, and Opus using engineer-written tests and estimated pass probabilities to inform model-selection policy.
- Containerized backend services with Docker and automated deployment through GitHub Actions CI/CD, using execution logs and agent traces to diagnose failures and maintain consistent release cadence.
- Enforced per-user token and spend controls with Slack alerts triggered at 50% and 80% of budget thresholds, restricting premium-model access to prevent cost overruns.
- Engineered PySpark feature pipelines to integrate historical warehouse records, brewery mappings, production activity, and calendar seasonality into consistent training and evaluation datasets for empties-return forecasting.
- Trained and evaluated six forecasting models — ARIMA, Prophet, XGBoost, LightGBM, LSTM, and Temporal Fusion Transformer — across multiple seasonal windows to identify per-horizon performance trade-offs.
- Built a neural meta-ensemble that learned dynamic weighting across six base models from historical residuals, producing next-week point forecasts for empty-bottle return volumes at the brewery level.
- Ran 10,000 Monte Carlo simulations sampling plausible future return volumes from model uncertainty and uncertain operational drivers, converting outcomes into P50, P90, and P95 probability distributions for planning.
- Applied Value-at-Risk and Conditional VaR to quantify adverse tail thresholds and expected severity beyond those thresholds, directly informing brewery bottle-availability risk assessments.
- Executed scenario analysis for individual operational driver changes and compound stress testing for simultaneous extreme events, translating results into cost-versus-risk trade-offs for capacity and replenishment decisions.
- Maintained weekly backtesting loops using WAPE, quantile coverage, and drift detection metrics; fed calibration findings back into model retraining, achieving reported forecast accuracy of 67%–91% across seasonal periods.
Education
Certifications
Powered by Drivetube · Create your own profile at drivetube.ai
Everything on Drivetube
Six products, two of them free forever. Start wherever you are.
Your free digital resume
at drivetube.ai/in/your-name: one true standard resume with a Hiring Snapshot (visa status, expected salary, notice period, work preference, relocation), an ATS-ready PDF download and a single shareable link. Free forever; interview requests come from verified employers and your contact details stay masked until you accept.
Documentation- Free Job BoardFree
Verified jobs, posted in the last 3 days
verified openings crawled ATS-by-ATS from 115,000+ real company career pages across 54 ATS platforms. Shows the true posting date from the source ATS — not when a listing was indexed — and deletes every general listing 3 days after it was actually posted. No ghost jobs, no ad-sponsored listings, no staffing reposts, no account needed.
Documentation - Resume Writing ServicesFrom $25.99
Written by senior career writers
human-written, ATS-optimised resumes by senior career writers, from ₹499.99 / $25.99. Available in every country, written to the destination country's own standard — a US resume, UK CV, German Lebenslauf and Indian resume are genuinely different documents. A paid service, separate from the free Drivetube Profile.
Documentation - Drivetube OneFrom $39.99 / 3 mo
Daily matched jobs, tailored resumes, or we apply for you
Drivetube's paid service. Community sends daily Jobsheets of fresh jobs matched to your profile on WhatsApp and email, with the Drivetube Community ($39.99 for 3 months or $69.99 for 6 months; ₹999 or ₹1,899 in India). Jobhunt is one-time application credits with Community included: one credit tailors one resume and its cover letter (100 for $99 with 3 months of Community, 250 for $199 with 6 months). Let Us Apply is an add-on to Jobhunt: a human expert applies 250 times for you on your credits ($0.60 each, $150). Everything lives on one dashboard at drivetube.ai/drivetubeone.
Documentation - Drivetube HireFree tier
For employers — no job postings, no applications
hiring with no job postings and no applications. Paste your real job description and AI matches it against candidates' true standard resumes, returning ranked candidates with a match %, matched and missing skills and written reasoning. Free tier included; employers pay, candidates never do.
Documentation
- Documentation
every product explained in full, with feature-by-feature comparisons against the job boards, AI apply tools, resume services and hiring platforms people actually use.
- Playbooks
tactical job-search guides. Each post is one named tactic with the situation it applies to, the exact steps, a copy-paste message and honest failure modes. Free, no account.
The job board covers the United States, India, United Kingdom, Canada, Europe and Australia, and Resume Writing Services are available in every country.
Pricing is identical on the web, iOS and Android — no app-store markup. Monthly Community membership is purchased in the iOS app only; the website sells annual plans. The Android app has no in-app purchase, so Android members subscribe on drivetube.ai and then sign in to the app with full access. Membership follows your account, not your device.