SlipstreamJobs tracks this role from the company's public career site. Apply directly on the employer's site.
Salary: USD 230,000 - 340,000 / annual
Baselayer is rebuilding the identity and trust layer for financial institutions across the United States. The company has built the most complete business graph in America by fusing public records, IRS data, sanctions lists, web signals, and fraud telemetry from 2,200+ financial institutions. With 98% match rates achieved in under two years (versus legacy credit bureaus at 60% over 50 years), Baselayer is trusted by over 20% of US financial institutions and is expanding into gig platforms, marketplaces, AI companies, and commerce infrastructure.
You will join a small, high-ownership team solving real-time entity resolution at scale—a graph AI problem, retrieval problem, and fraud-modeling problem combined. The team operates with minimal process overhead: ideas ship quickly, and the infrastructure you build becomes load-bearing for businesses.
In this role, you will own a slice of the agentic data enrichment surface end-to-end. Specifically, you will:
- Own industry and category classification of businesses from heterogeneous signals (name, website, directory presence, reviews).
- Build and maintain discovery and verification systems for a business's real web presence, filtering aggregators, parked domains, brand collisions, and impersonators.
- Link individuals to businesses via public web evidence (e.g., confirming named officers or employees genuinely work there).
- Develop risk and legitimacy scoring derived from web-presence signals, fed back into downstream underwriting.
- Build and evolve shared agent infrastructure: provider-agnostic base agents, shared toolset registry (browser navigation, search, scraping, structured database lookups, scoring), eval harness, and instrumentation surface for token-and-tool tracing.
- Own model selection, agent design, prompt and tool engineering, eval methodology, and cost control across your enrichment surface.
The company is at an inflection point: the graph is built, match rates are proven, and the hardest problems remain—graph embeddings, fraud propagation models across the business network, real-time traversal at sub-100ms latency, and expanding the identity layer beyond finance.
**Requirements:**
- Shipped LLM-driven agents to production (not notebooks or demos)—real users, real cost, real failure modes, real on-call experience.
- Strong async Python including structured-data libraries, modern web frameworks, and relational databases.
- Experience across multiple frontier LLM providers and at least one agent framework, with deep knowledge of failure modes.
- Built or maintained eval methodology: curated golden datasets, scoring functions, labeling guidelines, regression diagnostics.
- Browser automation experience: headless browsers, anti-bot evasion, authenticated flows.
- Informed opinions on structured-output reliability—when to use JSON-schema mode vs. function calling vs. extractor-on-top-of-text.
**Preferred experience:**
- Web scraping at scale: anti-bot evasion, residential proxies, request fingerprinting, authenticated flows, CDN defeats.
- Eval-framework experience (e.g., LangSmith, Braintrust, Evals, or custom).
- Entity resolution, record linkage, or fuzzy matching at scale.
- Browser-automation experience at the devtools-protocol level.
- Built a tool registry or toolset abstraction over multiple LLM providers.
- Cost and latency optimization: response caching, semantic caching, model routing (cheap-first then escalate), thinking-budget tuning, prompt-cache hit-rate work.
