Bridge-it is an early-stage EdTech company building a career-readiness platform for K-12 students, counselors, and districts. Students explore careers, colleges, and non-college pathways; build multi-year plans; and get guidance from an AI Copilot that is grounded in their own plans, school data, and a curated knowledge base. Counselors use the same platform to supervise, nudge, and intervene at scale.
The AI Copilot is the heart of the product — a production LangGraph-based agentic system serving real students, where being right matters more than being fluent. We hold ourselves to strict grounding, safety, and tenant-isolation standards, and we measure everything.
About the role
You will join a small AI team that owns the Copilot end to end: the agentic orchestration graph, retrieval and grounding, safety guardrails, response contracts with the frontend, and the evaluation platform that keeps it honest. We are mid-flight on an ambitious redesign — collapsing multiple experimental graph variants into a single workflow-first architecture with typed claim verification, constitutional rules, and per-turn decision records — and you would step directly into that work.
This is a senior role with a visible leadership runway. Beyond writing excellent code, we expect you to analyze ambiguous product requirements, turn them into well-scoped technical specs and tickets, distribute work across the team, and communicate trade-offs clearly to product and leadership. For the right person, this role grows into a techno-management position.
What you'll do
Design, build, and operate agentic LLM workflows in LangGraph — routing, planning, parallel retrieval, tool execution, verification, and streaming — serving live student and counselor traffic over WebSockets.
Own RAG quality: multi-source retrieval across vector stores (Milvus, pgvector), a knowledge graph (Neo4j), and structured platform data; chunking strategy; reranking; and grounded generation with typed claims and evidence citation.
Enforce safety and trust: input/output guardrails, constitutional rules, faithfulness verification, abstain/escalate behavior, and strict multi-tenant data isolation.
Build and extend our evaluation platform: online LLM-as-judge evaluators, offline regression suites on curated datasets (LangSmith), latency baselines, and A/B experimentation (PostHog feature flags).
Drive latency and cost optimization: model selection across providers (OpenAI, Gemini, open-weight models via OpenRouter), prompt and context engineering, selective retrieval, caching, and streaming intermediate responses.
Instrument and debug production behavior with LangSmith tracing, OpenTelemetry, New Relic, and Sentry; treat live-QA findings as first-class inputs to design.
Partner with backend (Django/Postgres), frontend (Next.js), and product teams on response envelopes, interaction contracts, and rollout plans.
Lead by doing: break epics into staged implementation plans with acceptance criteria, sequence work across engineers, review code, and mentor teammates.
What we're looking for (required)
4–8 years of professional software engineering experience, with at least 2–5 years building LLM-powered products in production (not just prototypes or notebooks) and 2 years of leading a team.
Strong Python engineering fundamentals — typing, testing, async, service design.
Strong grasp of agentic architecture patterns — supervisor/executor splits, planner–retriever–verifier pipelines, tool/capability registries, routing and fallback tiers — and sound judgment about when a workflow should be deterministic versus agentic.
Hands-on experience with agent orchestration frameworks (LangGraph strongly preferred; LangChain, or comparable state-machine/agent frameworks acceptable) including multi-step graphs, tool calling, and streaming.
Context engineering / harness engineering: deliberate construction of what the model sees each turn — context budgets, retrieval selection, system-prompt and tool-schema design, structured state passed between nodes — and the harness discipline (contracts, guards, fallbacks) that makes model behavior predictable.
Experience designing and optimizing agent memory: short-term conversation state (windowing, summarization/compaction, token-budget management) and long-term memory (user profiles, episodic stores, rollups) with sensible retrieval and staleness policies.
Deep practical knowledge of RAG systems: embedding models, vector databases (Milvus/Zilliz, pgvector, or similar), hybrid retrieval, rerankers (e.g., Cohere), chunking and knowledge-base design.
Proven experience building and operating LLM evaluation systems: online evaluators (LLM-as-judge) on live traffic, offline golden datasets and regression suites, faithfulness/hallucination and groundedness metrics, A/B experiment design — and the discipline to gate releases on eval results.
Hands-on AI observability: end-to-end trace instrumentation of agent runs (LangSmith, OpenTelemetry GenAI conventions, or similar), per-node latency/token/cost tracking, trajectory analysis, drift and quality-regression monitoring, and turning production traces into eval datasets and fixes.
Experience shipping structured outputs (JSON contracts, schema-validated responses) as the interface between models and product surfaces.
Excellent communication and interpersonal skills: you can explain a retrieval-quality trade-off to a product manager, write a spec another engineer can execute, and disagree constructively.
Demonstrated ability in requirement analysis and task decomposition — turning fuzzy asks into staged, dependency-ordered work — and willingness to step up into team-lead responsibilities.
Desirable
Experience at an early-stage startup — comfort with ambiguity, wearing multiple hats, and shipping under constraints.
Model Context Protocol (MCP): working knowledge of the spec and hands-on experience building or integrating MCP servers/clients to expose tools and data to agents.
Knowledge graphs (Neo4j/Cypher) in retrieval or recommendation contexts.
Durable workflow engines (Temporal) and event-driven architectures (Pub/Sub).
EdTech or other regulated/high-trust consumer domains.
Nice to have
AI infrastructure: model serving, GPU inference, gateway/router layers, and cost/latency-aware routing across providers.
SLMs and self-hosted LLMs: deploying and operating small language models or open-weight LLMs (vLLM, TGI, Ollama, or similar), quantization, fine-tuning/distillation for task-specific workloads, and knowing when a small model beats an API call.
Traditional ML: classification, ranking, embedding fine-tuning; scikit-learn/PyTorch.
Bridge-it is an equal opportunity employer. We welcome applicants of all backgrounds and are committed to an inclusive, respectful remote-first workplace.