Ishaan Chaturvedi
AI Engineer — Production LLM & Agentic Systems
London, UK · ishaan@nullsutra.com · LinkedIn · GitHub
Profile
AI Engineer specialising in production LLM and agentic systems, with 7+ years across enterprise machine learning and independent product building. I design, ship, and evaluate end-to-end AI inside regulated financial environments — self-hosted LLM deployment under GDPR, document-intelligence and agentic pipelines, enterprise-wide AI use-case discovery — focused on the parts that decide whether AI products survive production: evaluation, observability, structured outputs, and cost control. Independently, I architect and operate Alter, a self-hosted production AI “second brain,” and design generative-media tooling (Spiral).
Experience
PRA Group · London, UK Oct 2022 – Present
AI Engineer — Global AI Tiger Team 2025 – Present
- Selected onto PRA’s cross-organisation AI Tiger Team (reporting to an SVP) to identify and ship LLM and agentic AI use cases across US and international operations on an Azure stack.
- Deployed a self-hosted, multi-model LLM stack on 4× NVIDIA Tesla T4 GPUs to keep regulated data in-environment under GDPR — routing to purpose-specific models (fine-tuned Mistral for summarisation, DeepSeek-R1-Distill-Qwen for reasoning).
- Led a company-wide AI discovery initiative across every department, surfacing 15 high-value use cases, each quantified with effort–value scoring.
- Built an LLM + OCR pipeline over court documents extracting and classifying legal failure reasons into a tiered taxonomy.
- Developed agentic LLM pipelines for affluence detection and full customer-journey summarisation for frontline staff.
Data Scientist — Decision Science & AI Oct 2022 – 2025
- Portfolio valuation & forecasting: end-to-end pricing of non-performing loan portfolios (KNN peer selection, pay-curve generation, calibrated recovery models) — supporting 5 of 8 acquisitions now performing at ~98% of projected value.
- Monte Carlo simulation: state-based simulation with a classifier for account-state transitions and a regressor for payment amounts; benchmarked XGBoost, Random Forest, and a Time-Series Transformer.
- Collections decision science: WoE scorecards (avg AUC-ROC 0.74), 8 hierarchical customer clusters, optimised dialler contact limits.
ARCON TechSolutions · Mumbai, India Jan 2018 – Aug 2021
Data Scientist
- Built near-real-time anomaly-detection models for a privileged-access-management product.
- Developed an online hierarchical-clustering system for user-behaviour analytics.
- Led a two-person team to ship fraud-detection analytics tooling, later adopted across the product suite.
Selected Projects
Alter — Production AI “Second Brain” Lead / Primary Author
- Self-hosted, production AI “second brain” on a TypeScript / NestJS microservice architecture; an LLM recognition-and-orchestration layer classifies input, routes across five interaction modes, and auto-compiles a per-user knowledge wiki with semantic search.
- Agentic harness: multi-step, multi-mode workflows with automatic re-grounding, branching, and end-of-thread synthesis.
- Evals: golden eval suites as CLI runners with per-run cost budgets — prompts as tested, versioned components.
- Multi-provider + structured output: Anthropic / OpenAI / Google / Perplexity abstraction with complexity-based cost routing, prompt caching, and strict tool-use + Zod schemas with retry-repair.
- RAG: pgvector-backed hybrid (semantic + keyword) search over an event-driven reindex pipeline.
- Production-grade: encryption-by-default, CI-enforced privacy invariants, per-user cost caps, a full specification / ADR process.
Spiral — AI Filmmaking Tool Product & Design Engineering
- 0-to-1 product design for AI-assisted filmmaking built around iterative revision loops rather than a linear prompt-to-video pipeline.
- Agentic core: a typed dependency graph with three cascade modes (auto-propagate, flag-for-review, hard-invalidate) and a “converge” synthesis step.
- Consistency “locks” (style, character, motif, format, voice) auto-injected into downstream prompts; a clean free/paid generation boundary.
- Built through 7 interactive prototype iterations; validated by producing a complete short film inside it.
Generative Media Pipelines
- Directed and produced 2 AI short films end-to-end (story → moodboard → prompt engineering → consistent keyframes → ElevenLabs original scores); built an automated content pipeline.
Skills & Tools
- AI & LLM
- Multi-provider orchestration · agentic harness · structured outputs (Zod, tool-use) · prompt engineering, evals & cost budgeting · RAG / pgvector · fine-tuning & self-hosting · OCR
- Backend & Systems
- TypeScript · NestJS · Python · PostgreSQL · Redis / Bull · microservices & event-driven architecture · SSE streaming · ADRs
- Machine Learning
- XGBoost · Random Forest · KNN · hierarchical clustering · Monte Carlo simulation · WoE scorecards · time series · anomaly detection
- MLOps & Cloud
- Azure (AI Foundry, Fabric) · MLflow · Docker · GitLab CI · pytest / Jest / Playwright
- Generative AI
- AI video generation · AI music composition · automated content pipelines · prompt & storyboard design
- Regulatory
- GDPR · EU AI Act (regulated-finance AI delivery)
Education
MSc Artificial Intelligence (Distinction) · Queen Mary University of London 2021 – 2022
Dissertation: long-term planning agent for sparse-reward environments using Monte Carlo Graph Search & Quality-Diversity search — directly relevant to modern agentic systems.
BTech Computer Science & Engineering (Distinction) · Vellore Institute of Technology 2013 – 2017