Machine Learning Engineer

Maxwell Bond; Greater London; On-site; GBP 150,000 - 175,000; Full time

Job summary

Descrial, a startup within a twenty-year-old company, seeks a senior ML engineer to own core LangGraph/LangChain workflows in its London team. You will work hands-on on UK-based systems using Python, FastAPI services, and vector retrieval to drive investment-document analysis with strict cost controls and governance.

You will collaborate with the Head of Engineering, integrate major model providers, and craft scalable, auditable ML solutions for production in a high-stakes environment.

Qualifications

• 5+ years in software engineering with strong Python
• 2+ years building LLM/ML systems in production
• Hands-on agent/orchestration experience
• Production RAG with embeddings and vector databases
• Graph-backed retrieval experience with Neo4j and Cypher
• Experience with AI coding tools (e.g., Cursor, Claude Code)
• Degree in engineering or CS or equivalent experience

Responsibilities

• Build LangGraph/LangChain workflows with state, checkpoints, tool calls and human approval patterns
• Manage context windows, cost, and auditability in long-running runs
• Integrate model providers (Anthropic, OpenAI, Bedrock) and build fallbacks
• Develop and extend RAG pipelines with vector stores and graph databases
• Implement evaluation harnesses and data-privacy guardrails in the UK/GDPR context

Skills

Python

LLM systems

Agent orchestration

Graph retrieval

Evaluation harness

Education

Engineering degree

Computer Science degree

Tools

LangGraph

LangChain

Neo4j

Cypher

Job description

Machine Learning Engineer

King's Crosss

£150,000 - £175,000

We are hiring four people at senior and lead levels. Senior engineers own major capabilities, lead engineers also set technical direction and mentor others.

About Descrial

Descrial is a startup inside a twenty-year-old company

TresVista has run the analytical work behind private capital since 2006. 2,000 people, 400+ clients including some of the biggest names in the industry, no outside capital and no outside board. That was the first chapter. And we enjoyed getting here!

The next chapter is harder and more interesting

We are moving from services to software, from one firm's workflows to an industry's, and from executing investment decisions to shaping how they get made. Thirty of us in King's Cross, small and senior, with day one founder intensity.

At its fullest expression, Descrial changes how every serious investment firm decides.

The role

You report to the Head of Engineering and work hands-on on the UK core team. Our environment spans LangGraph and LangChain, leading model providers, Python/FastAPI services, and vector and graph retrieval. Token cost, latency and auditability are engineering constraints throughout.

Your remit

The agents and model layer

You build LangGraph/LangChain workflows with state, checkpoints, tool calls and human approval patterns. Pydantic-validated prompts and structured outputs support extraction, summarisation, classification and reasoning over investment documents. You integrate Anthropic, OpenAI and Amazon Bedrock, choose models by task and build fallbacks that absorb outages and provider changes. You manage context windows and compaction across long runs to control cost, and build and consume MCP tool servers for safe access to services, search and documents.

The retrieval

You own parsing, chunking, embeddings, Qdrant vector retrieval, reranking and hybrid search in RAG pipelines. You extend GraphRAG with Neo4j, Cypher and graph data modelling. Relevance evaluations guide improvements while indexed content stays isolated across tenants.

The evaluation and operation

You build golden datasets, LLM-as-judge evaluations and regression gates for prompt and model changes. LangSmith/LangFuse-class instrumentation makes latency, cost and quality visible for every model call. You apply tool-use constraints, output validation and data-privacy guardrails in the UK/GDPR context. You ship Python 3.11+/FastAPI services with PostgreSQL and operate on AWS through CI/CD quality gates.

Your experience

• You have 5+ years in software engineering with strong Python, including 2+ years building LLM or ML systems used in production.
• You have deep hands-on agent/orchestration experience with LangGraph, LangChain or comparable frameworks beyond prototypes.
• You have production RAG experience with embeddings, vector databases, chunking and retrieval evaluation.
• You have graph-backed retrieval experience or graph-modelling foundations that let you learn Neo4j and Cypher quickly.
• You can describe an evaluation harness you built and how it governed changes.
• You use AI coding tools daily, such as Cursor or Claude Code, and have a considered view of where they help.
• You have a degree in engineering, computer science or a related field, or equivalent experience from a strong technical background.

Useful to have, none required

• Qdrant, Weaviate or pgvector in production, and ingestion tools such as Docling or unstructured
• MCP server development and observability tools such as LangSmith, LangFuse or Ragas
• Fine-tuning, embedding-model selection or open-weight deployment
• Financial documents such as filings, fund documents and research

The reality

This will be difficult.

It demands pace and accuracy. We are building autonomous systems inside high-stakes live businesses, where decisions carry real consequences and impact and waiting for everyone to agree is not always an option.

The systems will sometimes get things wrong. In this phase the London team needs to create the evaluations, visibilities and safeguards that catch mistakes early, reducing them over time and stopping the same failures happening again.

If you want to build the foundations of an enterprise grade technology platform to service a multi-billion $ serviceable market opportunity, and share in that growth, this is your chance to join us.

In our conversations

Come ready to discuss three pieces of work:

• An agent or retrieval system you shipped: the architecture and how it performed.
• An evaluation harness you built: the failures it caught and the release decisions it changed.
• A model, context or routing decision that improved quality, latency or cost.

Compensation

The all-inclusive package is typically £150,000–£175,000 a year, depending on experience.

Practical details

• King's Cross, London. In office (3 to 5 days a week), with colleagues in London and Bengaluru.
• Visa sponsorship is not available for this role.
• We aim to finish the process within three weeks. Tell us if you need an adjustment to the interviews.
• Descrial is an equal opportunity employer.