Krishna Chaitanya
Makthala

AI Engineer — Multi-Agent Systems & Production LLMs

An AI Engineer focused on systems that move from architecture to production — multi-agent platforms, RAG pipelines, and ML forecasting. Based in Bristol, UK, with a background in enterprise software delivery.

View Projects Download CV GitHub LinkedIn
Available now 📍 Bristol, UK 🎓 MSc Data Science, UWE Bristol 🇬🇧 Right to work until Dec 2027

Background

Professional Background

I am an AI Engineer with a foundation in enterprise software delivery and a focus on production-grade systems.

The career path began with two years of enterprise software delivery at Accenture — production engineering across the full lifecycle, working within SLA frameworks and cross-functional teams. That grounding in how software ships in real organisations shapes the approach to AI engineering: models are one component of a larger system, not the whole product.

The Avsar contract was a six-month engagement taking a multi-agent AI platform from initial architecture to live production on AWS. What it demonstrated at a higher level was how software engineering discipline — testing, deployment, monitoring — applies directly to AI systems in the same way it applies to any production service. The technical detail is in the Projects section.

The MSc capstone at UWE Bristol took a different angle — a comparative evaluation of four forecasting architectures across 12 product categories on a held-out test set, with explicit statistical baselines. 2.75% MAPE against a 6.8% baseline.

2.75%
MAPE — retail demand forecasting (vs 6.8% baseline)
3
Specialist AI agents deployed on AWS ECS Fargate
<1s
End-to-end response — AskMyDoc (FAISS retrieval + LLM generation)

Projects

Projects

Production deployments, open-source tooling, and applied ML research — spanning multi-agent systems, retrieval-augmented generation, and forecasting.

Commercial · NDA
Avsar
Multi-Agent AI Platform
✓ 3 agents on AWS ECS Fargate
Multi-agent AI platform built for hotel operations on AWS. Three specialist agents — Concierge, Booking Assistant, and Complaint Triage — each with dedicated tool sets and LangChain orchestration, sharing a Pinecone knowledge base built on a custom ingestion pipeline: document chunking, text-embedding-3-small vectorisation, and semantic retrieval. Containerised with Docker, deployed to ECS Fargate, with automated CI/CD via GitHub Actions.

This engagement clarified what separates AI that demos cleanly from AI that holds up in production — working through failure modes in retrieval, infrastructure constraints under real load, monitoring requirements, and the discipline of writing pipelines that remain stable across diverse and unpredictable user inputs.
LangChainOpenAIPinecone FastAPIDockerAWS ECS GitHub Actions
GitHub ↗
AskMyDoc
RAG Document Q&A
✓ Sub-1s standard · under 4s on 200-page files
AskMyDoc is a document question-answering system built around a specific failure mode in retrieval-augmented generation: when a query is ambiguous, standard RAG retrieves plausible but incorrect context and the LLM generates a confident wrong answer. AskMyDoc introduces a Clarification Gate — an LLM-powered pre-retrieval layer that evaluates query specificity before retrieval begins. If the query is too broad or ambiguous, the system surfaces a targeted follow-up question rather than proceeding with unreliable retrieval. When the query is sufficiently specific, it passes directly to FAISS vector search, followed by LLM generation — end to end in under one second. Open source on GitHub.
LangChainOpenAIFAISS StreamlitPython
MSc Capstone
Retail Demand Forecasting
4 architectures · 12 categories
✓ 2.75% MAPE vs 6.8% baseline
Comparative evaluation of four forecasting architectures — Prophet, ARIMA, XGBoost, and LSTM — across 12 COICOP product categories on a held-out test set, with explicit statistical baselines. 2.75% MAPE overall — Prophet performed best in 10 of 12 categories. Best individual performance in Recreation (0.19%) and Transport (0.23%). Deployed as an interactive Streamlit application with configurable forecast horizons and per-category model selection.
ProphetXGBoostARIMA LSTMStreamlitPython
Avsar — System Architecture Commercial · NDA · Diagram sanitised
%%{init: {'theme': 'default', 'themeVariables': {'background': '#faf8f5', 'primaryColor': '#e6dfd5', 'primaryTextColor': '#1c1917', 'primaryBorderColor': '#1e3a8a', 'lineColor': '#57534e', 'secondaryColor': '#f2eee9', 'tertiaryColor': '#f2eee9', 'edgeLabelBackground': '#faf8f5', 'fontFamily': 'Inter', 'fontSize': '14px'}}}%% graph LR U([User]) --> F[FastAPI\nGateway] F --> R{Agent\nRouter} R -->|guest queries| CA[Concierge\nAgent] R -->|reservations| BA[Booking\nAgent] R -->|complaints| TA[Triage\nAgent] CA --> RAG[(Pinecone\nVector DB)] BA --> RAG TA --> RAG RAG --> LLM[OpenAI\nLLM] LLM --> Out([Response]) subgraph AWS[AWS Infrastructure] F R CA BA TA end

Experience

Professional Experience

Contract AI engineering followed by two years of enterprise software delivery.

Dec 2025
— Jun 2026
AI Engineer (Contract)
Avsar  ·  Remote, UK
  • Designed and deployed a multi-agent AI platform on AWS ECS Fargate — three specialist agents (Concierge, Booking, Complaint Triage) with LangChain orchestration and dedicated tool sets per agent.
  • Built a shared RAG knowledge base using Pinecone and OpenAI embeddings, with a custom document chunking and ingestion pipeline serving all three agents from a single retrieval layer.
  • Delivered a FastAPI backend, Docker containerisation, and GitHub Actions CI/CD pipeline — automated build, test, and deployment to AWS ECR and ECS Fargate on every push.
Sep 2021
— Sep 2023
Application Development Associate
Accenture  ·  Hyderabad, India
  • Engineered Python automation scripts to streamline multi-environment deployment workflows, reducing manual release effort and improving delivery-cycle efficiency across production environments.
  • Built and maintained ASP.NET MVC web applications and PowerShell configuration-management workflows, improving system uptime by 20% during peak retail demand periods.
  • Analysed operational and retail datasets using SQL to identify incident patterns, supporting data-driven decisions and contributing to 100% SLA compliance — collaborating cross-functionally on root-cause analysis and production issue resolution.

Capabilities

What I Bring to the Business

Core areas of expertise, with examples from delivered work.

Production AI Engineering
The Avsar engagement covered the full path from system design to live deployment on AWS - multi-agent orchestration, shared RAG infrastructure, containerised services on ECS Fargate, and an automated CI/CD pipeline. Production AI engineering means the complete stack, not just the model layer.
LangChain · FastAPI · Docker · AWS ECS Fargate · ECR · GitHub Actions · S3
Retrieval-Augmented Generation
Building effective RAG systems goes beyond connecting a vector database to an LLM. The AskMyDoc project introduced a pre-retrieval Clarification Gate that resolves query ambiguity before the retrieval step, preventing retrieval from proceeding on underspecified queries. The Avsar platform uses a shared Pinecone knowledge base with custom chunking and text-embedding-3-small semantic search serving three agents from one retrieval layer.
Pinecone · FAISS · OpenAI Embeddings · text-embedding-3-small · LangChain
ML and Forecasting
The MSc capstone was a structured comparative evaluation - four model architectures, 12 COICOP product categories, held-out test set, explicit statistical baselines. The methodology follows standard time-series validation practices; the 2.75% MAPE vs 6.8% baseline result is grounded in that process, with every comparison documented against a proper held-out evaluation.
Prophet · XGBoost · ARIMA · LSTM · scikit-learn · pandas · NumPy · SQL
Enterprise Delivery Background
Two years at Accenture provided grounding in enterprise software delivery - SLA frameworks, cross-functional teamwork, and production release cycles. That context shapes how AI systems are approached: as components of larger engineering systems, with the disciplines of testing, documentation, and stakeholder communication that entails.
Python · SQL · AWS Cloud Practitioner · Azure AZ-900 · Agile delivery

Education

Education

MSc Data Science
University of the West of England, Bristol
Sep 2024 — Sep 2025
The MSc changed how data problems are approached — particularly around evaluation honesty. Working through statistical learning, Bayesian reasoning, and time-series modelling built a more disciplined instinct for when a model is actually working versus when it just looks good on training data. The programme pushed hard on held-out evaluation, explicit baselines, and reproducibility — habits that carry directly into production AI work.

The capstone was where that rigour was tested in practice. Four forecasting architectures, 12 product categories, results that had to be defensible against a proper statistical baseline. Arriving at 2.75% MAPE against a 6.8% baseline was satisfying not because of the number, but because the methodology behind it was sound.
BTech Electronics & Communications Engineering
SNIST, Hyderabad, India
Aug 2017 — Aug 2021
Four-year undergraduate degree in Electronics and Communications Engineering. Provided foundations in mathematics, signal processing, and systems thinking that underpin the machine learning and engineering work that followed.

Contact

Let's Work Together

Available for AI Engineer and GenAI roles in the UK. Immediately available, with full right to work until December 2027. Happy to discuss opportunities or answer questions about any of the work above.

Email LinkedIn GitHub