Hardware startup — RAG over internal runbooks
MQTT field data → Cloud Storage → chunked runbook corpus → pgvector → OpenAI function-calling agent. 62% of support tickets auto-resolved.
AI & agentic systems that survive contact with production.
Custom RAG pipelines, LLM evals, voice agents, Copilot extensions, and MLOps that doesn’t break when a model version changes. We build the boring parts — retries, fallbacks, evals, observability — that turn a demo into a product.
Rate $80 – $100 USD / hour · fixed-fee and retainer available
What you get
- RAG pipeline
- with the boring parts done — chunking strategy, embeddings, vector store, reranking, observability, cost dashboards.
- Eval harness
- against a frozen labeled set, with regression alerts on every prompt or model change.
- Agent orchestration
- with typed tool definitions, retry policies, human-in-the-loop checkpoints, and full trace logging.
- Microsoft 365 Copilot
- declarative agents, message-extension plugins, and Graph connectors — including the tenant-side security review.
- MLOps
- on Vertex AI / SageMaker / Kubeflow — model registry, CI for models, A/B routing, canary releases.
- Cost controls
- per-tenant budget guards, model-tier routing, prompt caching, and a monthly cost review.
How we engage
- 01
Discovery (2 weeks, fixed fee)
We read your data, your existing prompts, and your failure modes. We write a one-page “what good looks like” and a cost model.
- 02
Build (4–8 weeks)
RAG pipeline, eval harness, and the first production agent. We ship behind a feature flag.
- 03
Operate (ongoing)
Optional retainer: weekly model-eval runs, monthly cost/perf review, on-call for production incidents.
Stack we work in
LLM providers
OpenAI, Anthropic, Gemini, Mistral, vLLM, Ollama, Vertex AI
Frameworks
LangGraph, LlamaIndex, DSPy, Haystack, custom orchestration
Vector stores
Pinecone, Weaviate, pgvector, Qdrant, Vertex AI Vector Search
Evals
Promptfoo, Braintrust, custom labeled sets, golden traces
Voice
Twilio + Deepgram + Cartesia, Vapi, custom pipelines
Copilot / M365
Declarative agents, Teams message extensions, Graph connectors
Reference architectures
Anonymized patterns from real engagements. Client names omitted; details available under NDA.
Copilot extension for a legal-tech SaaS
Declarative agent + Graph connector over a private DMS, deployed to a regulated tenant. Passed the Microsoft security review on first submission.
Voice agent for last-mile logistics
Bilingual (English + Hindi) voice agent handling 1,200 calls/day. p95 latency < 1.4s. Cost per call < $0.12.
Questions we get asked
How do you stop LLM demos from breaking in production?
Evals against a frozen golden set, model-version pinning with explicit upgrade windows, output-schema validation, and observability on cost and latency per call.
Which model providers do you work with?
OpenAI, Anthropic, Google Gemini, Mistral, open-source via vLLM/Ollama, and on-prem via Vertex AI. We are not a reseller — we recommend per workload.
Can you build Microsoft 365 Copilot extensions for our tenant?
Yes. Declarative agents, message-extension plugins, and Graph connectors for the Microsoft 365 ecosystem, including the security review process.
What does a RAG engagement look like end to end?
Document audit → chunking + embedding strategy → vector store → retrieval + reranking → evals on a labeled set → deploy with usage telemetry.
Do you do fine-tuning?
When it actually helps. Most startups do not need fine-tuning — good retrieval, good prompts, and a long enough context window get you 90% of the way. We tell you honestly when fine-tuning is the right call.
Let’s scope it properly.
A 30-minute call. No deck, no pitch — we read your repo or your architecture diagram and tell you what’s realistic.