AI Discovery Agent Developer - Part-time - Worldwide

Objective

Create a production-ready Discovery Agent that interviews prospects, understands answers in context, asks clarifying questions when inputs are incomplete, ingests public company data (website + LinkedIn) and uploaded docs, then outputs a structured discovery report and solution recommendations. Must support multiple LLMs (OpenAI, Anthropic, Llama via AWS Bedrock, etc.).

Target Users

Logistics SMBs (10–100 employees). Agent should be industry-aware but generalizable to other verticals later.

Preferred Stack

• AWS: API Gateway or ALB + Lambda (or Fargate), S3, CloudFront, OpenSearch Serverless (or pgvector), SQS/EventBridge, Secrets Manager, CloudWatch.
• Langs: TypeScript (Node) or Python (FastAPI).
• LLMs: OpenAI, Anthropic, Llama (Bedrock) via a pluggable adapter.
• RAG: URL/PDF ingestion, chunking/embeddings, retrieval with citations.
• IaC & CI/CD: Terraform + GitHub Actions.

Scope & Deliverables

1. **Conversation API & Store
   **Endpoints: POST /sessions, POST /sessions/{id}/messages, GET /sessions/{id}/summary|report.
   Session persistence, consent flags, artifact refs, rate limits.
2. **Intelligent Clarification
   **Detect gaps/contradictions vs. a Discovery JSON schema; ask targeted follow-ups with configurable limits/timeouts.
3. **Ingestion & Retrieval
   **Crawl provided website (1–3 levels, robots-aware), accept LinkedIn URL(s), handle PDF/DOCX uploads.
   Embeddings + vector store (OpenSearch Serverless preferred).
   Retrieval tuned for concise answers + evidence snippets.
4. **Multi-LLM Adapter
   **Providers: OpenAI, Anthropic, Bedrock (Llama, etc.).
   Simple routing by task/cost/latency; streaming responses (SSE).
5. Outputs
   Discovery JSON: company profile, systems/data sources, workflows, pain points, volumes/SLAs, compliance, integration priorities.
   Human-readable summary (Markdown/PDF) and Recommendation bundle (candidate solutions with pros/cons + T-shirt size).
6. **Admin Insights (MVP)
   **Metrics: completion rate, # clarifications, retrieval hit rate, model spend estimate; simple ROI stub.
7. **Security & Guardrails
   **Keys in Secrets Manager, PII redaction toggle, domain allowlist for crawlers, prompt-injection filters, redacted logs.
8. **Infrastructure & DevEx
   **Terraform modules for all resources; GitHub Actions pipeline; CloudWatch logs/metrics.
9. **Docs & Handoff
   **README, runbooks, architecture diagram, threat-model checklist, test plan; admin how-to for prompts/router policies.

Non-Functional Requirements

• Perf: P95 chat turn < 3s (with retrieval); ingestion jobs typically < 5 min.
• Cost: Serverless first; surface per-session inference/infra spend.
• Reliability: Timeouts/retries; DLQ for failed ingestions.
• Privacy: No training on client data; region us-east-1 unless specified.

Milestones (example)

• M0 (1 wk): Architecture + Terraform skeleton + Hello-World API.
• M1 (1–2 wks): Chat + clarification loop + multi-LLM adapter.
• M2 (1–2 wks): Ingestion (web/LinkedIn/PDF) + embeddings + retrieval.
• M3 (1 wk): Discovery JSON + summary/PDF + admin metrics stub.
• M4 (0.5–1 wk): Guardrails, tests, docs, final demo.

Acceptance (MVP)

• Full session flow returns valid Discovery JSON and downloadable PDF summary.
• Evidence snippets demonstrate that ingestion informs answers.
• Metrics endpoints return non-zero values after test runs.
• Terraform can deploy all required resources into our AWS account.

Candidate Requirements

• Proven delivery of agentic or RAG systems in production.
• Strong AWS (Lambda/Fargate, API Gateway/ALB, S3/CloudFront, Secrets Manager, CloudWatch).
• TypeScript/Node or Python expertise.
• Experience with OpenAI/Anthropic/Bedrock; embeddings/vector DBs.
• Terraform + GitHub Actions.Nice-to-have: OpenSearch tuning; prompt-injection defenses; LinkedIn/site ingestion; VAPI/Voice/Twilio.