What a Week Looks Like in This Role
You will build agents that run the CFO's function — accounts payable, month-end close and
record-to-report, reconciliation, vendor and compliance workflows — end to end, inside enterprise ERP environments. These systems carry out finance processes rather than just suggest actions: they interpret source documents, apply accounting and control logic, perform transactions in SAP, and escalate only genuine exceptions for human review.
You will own outcomes end to end rather than individual components. Performance is measured through finance outcomes: touchless processing rate, exceptions per thousand documents, close cycle time and reduction in manual controller effort.
Key Responsibilities
Agents that run finance workflows
- Own end-to-end finance workflows as autonomous agents: invoice-to-pay, three-way matching and exception handling, GL close tasks, reconciliation, and vendor and compliance checks.
- Design tool and action surfaces over ERP systems, control flow, error recovery, human-in-the-loop checkpoints, approval and segregation-of-duty guardrails, and full auditability of every action taken on financial data.
- Build against production enterprise surfaces — SAP APIs, ERP user interfaces via computer-use agents, documents and file-based integrations — using LangGraph, LangChain or custom orchestration.
Memory and the finance knowledge graph
- Own agent memory: context management (windowing, summarisation, compaction), long-term memory, retrieval and forgetting policies, and cross-cycle state for long-running agents.
- Model the finance domain as a knowledge graph — vendors, invoices, POs, GRNs, GL accounts, contracts, approvals and their provenance — in a graph database (Neo4j or equivalent) alongside vector search, so every output can be substantiated for finance and audit review.
Evaluation and model strategy
- Build and extend the evaluation platform: golden datasets, workflow-level metrics, regression gates in CI, automated judging calibrated against human review, and drift detection on live traffic. No change ships without a measured evaluation delta.
- Work across model families (Anthropic, OpenAI, Google, open-weight) and own routing — which model handles which step at what quality, cost and latency point, with tiering, cascading and provider failover.
Scale and platform
- Own latency, throughput and unit economics for agents and pipelines running at large document and transaction volumes, maintaining p50/p95/p99 latency and cost-per-workflow targets as the customer base grows.
- Maintain the API surface (FastAPI, REST) that delivers these capabilities into the products, and mentor engineers working alongside AI systems.
Required Experience
- Agents in production: delivered a multi-step agent that performs actions in a production system, with a clear understanding of its failure modes and the controls used to contain them.
- Memory: designed the memory and context strategy for a long-running agent, including retention, summarisation and pruning decisions.
- Evaluation: hands-on experience building an evaluation harness, including dataset construction, metric selection and calibration of automated judging.
- Multiple models: delivered systems on at least two model providers or families, with experience migrating a live workload between them.
- Scale: operated an LLM or ML system in production at significant request volume, with measurable before-and-after results.
Qualifications
- Degree in Computer Science, Computer Engineering or a related field.
- Advanced proficiency in Python and modern AI/ML development practices.
- Hands-on experience with agent orchestration frameworks (LangGraph, LangChain, OpenAI Agents SDK), or with custom orchestration built directly on model APIs.
- Working knowledge of graph data modelling and a graph database (Neo4j, Neptune, ArangoDB): schema design, query (Cypher or Gremlin), and judgement on when a graph beats a relational or vector store.
- Strong API and data engineering fundamentals (FastAPI, PostgreSQL, vector stores) and cloud-native deployment on GCP, AWS or Azure.
- Sound judgement on where LLM-based approaches fit and where deterministic methods are better.
Tech Stack
Python · LangGraph / LangChain · Anthropic, OpenAI and Google model APIs · FastAPI · PostgreSQL with pgvector · Neo4j · Elasticsearch · Java/Spring Boot and React · GCP and AWS · Docker · in-house evals platform and in-house OCR. Familiarity helps but is not a prerequisite.
Preferred / Good to Have
- Finance, accounting, audit or ERP (SAP) domain exposure, or a strong interest in building domain depth.
- Computer-use or browser automation agents; GraphRAG and ontology design; document AI / OCR; fine-tuning or serving open-weight models.
- LLM tracing and observability (LangSmith, Langfuse, OpenTelemetry); enterprise AI security, governance and compliance frameworks.
- Regular use of AI coding agents in day-to-day development.
Why This Role
Uniquely positioned: the client sits where Payments, Lending and SaaS meet — three of the
fastest-growing spaces globally and in India — and its agents run against live enterprise finance operations at national scale.
Wealth creation: competitive compensation, strong incentives and meaningful equity in a fast-growing early-stage company.
Direct impact and growth: tangible impact, mentoring and clear growth paths on both IC and management tracks.
Benefits:
- Health insurance
- Paid time off
- Provident Fund
Application Question(s):
- How many years of experience do you have building LLM-based systems that are running in production (not PoCs or demos)?
- Have you built and deployed a multi-step AI agent that performs actions in a live system (e.g., updates records, posts transactions, closes tickets), not just a chatbot or RAG Q&A?
- Have you built an evaluation setup for an LLM system, including a golden/test dataset and metrics used to decide whether a change ships?
- Are you currently based in Mumbai, or willing to relocate to Mumbai (Andheri East) and work from the office?
- Have you used a graph database (Neo4j, Neptune or ArangoDB) in a production project, including schema design and writing Cypher/Gremlin queries?
- Have you worked with finance, accounting or ERP (SAP) data such as invoices, purchase orders or GL entries?
- Have you built agents using LangGraph, LangChain, OpenAI Agents SDK, or custom orchestration on model APIs?
- Have you designed memory or context management (summarisation, compaction, long-term memory) for a long-running agent?
- Have you shipped production systems on at least two LLM providers (e.g., OpenAI, Anthropic, Google, open-weight) and migrated a live workload from one to another?
Work Location: In person
