Projects
Fifteen systems, built in dependency order.
Each phase reuses the infrastructure — auth, eval, observability, deploy — established in the phase before it. This is a build order, not a list.
1 in progress · 14 planned
Core foundations(4)RAG, agentic orchestration, document intelligence, MCP — the infrastructure every later project reuses.
- 01RAG
RAG Starter Kit
In progressSMBs (law firms, clinics, agencies) have policies, SOPs, and past client emails scattered across drives with no searchable interface.
Next.js chat UIFastAPILangGraph retrieval nodepgvector similarity searchLLM synthesis + citationsEvery call traced in Langfuse — a DeepEval CI check fails the build if retrieval precision drops below threshold.
- FastAPI
- Next.js
- Postgres + pgvector
- LangGraph
- LlamaIndex
How it works
- 01
Ingest
Documents (PDFs, DOCX, email exports) go through the FastAPI backend, get chunked with LlamaIndex, embedded with OpenAI's text-embedding-3-small, and stored in Postgres — each chunk's embedding lives in a pgvector Vector(1536) column next to the source document and page number it came from.
- 02
Ask
The Next.js chat UI sends the question to FastAPI, which hands it to a LangGraph retrieval node — a small state graph rather than a single hard-coded function, so the retrieval logic can grow branches later without a rewrite.
- 03
Retrieve
The node embeds the incoming question and runs a similarity search against pgvector to pull back the chunks most relevant to it.
- 04
Synthesize
Those chunks go to the LLM with instructions to answer only from what was retrieved and attach a citation — source document and page — to every claim, so nothing in the answer is unsourced.
- 05
Trace
Every step of that call — retrieval, synthesis, latency, token counts — is logged to Langfuse, so any single run can be pulled up and inspected after the fact.
- 06
Gate
Before any change ships, a DeepEval golden set of 20-30 real question/answer pairs runs against the retrieval pipeline in CI. If precision drops below threshold, the build fails and the deploy is blocked — the eval isn't a report, it's a gate.
- 02Agent
Multi-Tenant Customer Support Agent
PlannedSaaS companies are drowning in repetitive tickets and need escalation logic they can trust.
Ticket webhookClassifyRetrieveDraft replyConfidence gateAuto-resolveorEscalate to humanCRM write-back and a per-tenant Langfuse trace on every ticket.
- LangGraph
- Zendesk/Intercom API
- Stripe
- multi-tenant Postgres
- 03Extraction
Document Intelligence Pipeline
PlannedLegal, insurance, and real-estate teams manually re-read contracts for clause risk.
UploadParse (native / OCR fallback)Chunk vs. clause schemaRisk-flag scoringReviewer UIEvery flag links back to its exact page in the source document.
- unstructured.io / LlamaParse
- OCR fallback
- Pydantic-constrained generation
- 04MCP
MCP Server for Internal Data Access
PlannedEvery agent needs to query internal systems (CRM, database, ticketing) but hand-rolled tool schemas don't compose.
Claude Desktop / CursorMCP server (auth-scoped)list_customers · get_order_history · flag_risk_accountInternal Postgres / CRMPublished to the MCP registry — callable by any client, including the agents from projects 02 and 05.
- MCP Python/TS SDK
- OAuth-scoped access control
Agentic & multi-system(3)Multi-agent orchestration, browser automation, real-time voice.
- 05Multi-agent
Multi-Agent Workflow Automation
PlannedSales/ops teams manually qualify inbound leads and draft outreach.
CrewAI · LangGraph · enrichment APIs
- 06Browser agent
Browser Agent for Competitive Monitoring
PlannedBusinesses want to track competitor pricing/inventory changes without a human checking daily.
Playwright · browser-use-style agent loop · cron · diff-based change detection
- 07Voice
Voice AI Receptionist
PlannedLocal service businesses (clinics, salons, restaurants) miss calls constantly and can't afford 24/7 staff.
Whisper / streaming STT · ElevenLabs · Twilio · LangGraph booking flow
Deep technical differentiation(4)Fine-tuning, guardrails, graph retrieval, classical ML — proof this goes past API wrapping.
- 08Fine-tuning
Fine-Tuned Local Model + vLLM Deployment
PlannedAPI costs and latency don't scale for a narrow, high-volume classification or style-matching task.
PyTorch · LoRA / PEFT · vLLM · Ollama
- 09OSS
Open-Source Eval & Guardrails Toolkit
PlannedEval and injection-defense get rebuilt ad hoc inside every LLM project instead of shipped once as a reusable, testable layer.
Python package · PyPI · GitHub Actions
- 10GraphRAG
GraphRAG for Relationship Intelligence
PlannedSome questions aren't 'find the similar chunk,' they're 'how are these entities connected' — vector search alone fails here.
Neo4j · entity/relationship extraction · hybrid retrieval
- 11Classical ML
Recommendation & Personalization Engine
PlannedPersonalized ranking rarely needs an LLM in the loop — embeddings and collaborative filtering are usually faster and cheaper to serve.
scikit-learn · embedding-based similarity · Redis
Operational maturity & monetization(4)Orchestration across everything above, a shipped product, an audit layer, a public asset.
- 12MLOps
MLOps Orchestration Layer
PlannedAll prior projects need a real retraining/monitoring loop, not manual reruns.
Airflow / Prefect · MLflow · Terraform · OpenTelemetry
- 13SaaS
Monetized Micro-SaaS (Support Agent, Productized)
PlannedTake the multi-tenant support agent and turn it into a product with paying customers, not a demo.
Stripe billing · marketing landing page · self-serve onboarding
- 14Security
Enterprise Security/Compliance Layer for Agents
PlannedEnterprises won't deploy an agent that can't prove it's safe.
structured audit logging · RBAC middleware · approval-queue UI
- 15Benchmark
Public LLM/Vector-DB Cost & Latency Benchmark Dashboard
PlannedNobody has a trustworthy, independent, continuously-updated comparison of LLM providers and vector DBs on a fixed task.
Airflow · Next.js · vLLM · pgvector / Qdrant