Advanced RAG
Hybrid search (BM25 + dense), cross-encoder re-ranking, query rewriting, document-aware chunking, citation tracking, metadata filtering, multi-tenant retrieval.
Q3 2026 — 1 engagement slot open
Production AI engineering, without the partner-shop markup.
The same caliber of work you’d get from an Anthropic or OpenAI implementation partner — advanced RAG, Graph RAG, fine-tuning, multi-step agents — at boutique pricing and without the enterprise sales process.
Fully async. No discovery calls, no demos, no zoom marathons. You send a brief → we send a written proposal in 48 hours → you decide.
- 20+
- Years in practice
- 68
- Systems shipped
- 14
- LLM apps in prod
- $35k
- Engagements from
PositioningPartner-shop expertise.
Partner-shop expertise.
Boutique-shop pricing.
Anthropic and OpenAI implementation partners deliver excellent work. They also charge $150k+ for engagements that start with a six-week sales process. We deliver the same caliber of production AI work, fully async, at one-third the price.
| Partner shops | Mini Trends | |
|---|---|---|
| Engagement floor | $150k – $500k | $35k |
| Sales process | 4 – 8 weeks of calls | 48-hour written proposal |
| Senior engineer on the keyboard | Sometimes (juniors hidden in rate) | Always (senior-led, no juniors) |
| AI capability ceiling | Frontier (RAG, agents, fine-tuning) | Frontier (RAG, Graph RAG, fine-tuning, agents) |
| Code & infra ownership | Yours, eventually | Yours, day one |
| Async delivery | No | Yes — entire engagement |
Mini Trends is independent. We are not affiliated with Anthropic or OpenAI; the comparison is on capability and price.
AI capabilities
Anything AI. Done at the production bar.
The full surface area of modern AI engineering, delivered by a senior-led team. If it is shipping in production somewhere in 2026, we have shipped it.
Graph RAG
Knowledge graph construction over unstructured data (LangChain, LlamaIndex, custom). Entity extraction, relationship modeling, graph-aware retrieval for multi-hop reasoning.
Fine-tuning & adaptation
LoRA, QLoRA, and full fine-tunes on Llama, Mistral, Qwen, and open-weight models. Distillation from frontier models to smaller production-targets. DPO and preference-tuning where it pays.
AI agents
Single-agent and multi-agent systems with tool use, function calling, durable state (Temporal, Inngest), human-in-the-loop checkpoints, and replayable execution.
Evals & observability
Eval harnesses with LLM-as-judge and deterministic scoring. Production tracing with LangSmith, Braintrust, Phoenix. Continuous shadow evals on real traffic. Drift detection.
Multi-modal
Vision (Claude, GPT-4V, Gemini), document understanding, OCR pipelines, audio (Whisper, Deepgram), and voice agents (LiveKit, Pipecat, Vapi).
Embeddings & vector infra
Vector database selection and tuning (pgvector, Qdrant, Pinecone, Weaviate, Turbopuffer). Embedding model selection. Hybrid index design at scale.
Cost & latency optimization
Prompt caching, model routing, speculative decoding, structured output for reliability, response streaming, batch APIs. Bills cut 40–80% on most engagements.
Frontier-model integration
Claude (Sonnet 4.6, Opus 4.7), GPT-5, Gemini 2.5 Pro, Llama 4. Provider-agnostic abstractions where it pays; provider-specific features where it does not.
Cloud & infrastructure
We know the clouds, deeply.
AI is half the job. The other half is the infrastructure that runs it. We are equally fluent in AWS, GCP, Azure, and the modern PaaS layer — and equally opinionated about when each is the right call.
AWS — production-deep
Bedrock, ECS Fargate, Lambda, RDS Aurora, S3, EventBridge, Step Functions, OpenSearch, IAM. Well-Architected reviews. Cost optimization on six- and seven-figure monthly bills.
Google Cloud
Vertex AI, Cloud Run, BigQuery, Pub/Sub, Cloud SQL, Firestore, Cloud Functions, Workflows, IAM. Multi-region deployments and BigQuery ML pipelines.
Azure
Azure OpenAI Service, Container Apps, Cosmos DB, Functions, Service Bus, AKS, Key Vault. Comfortable when the enterprise has already standardized here.
Modern PaaS
Vercel, Fly.io, Railway, Render, Cloudflare Workers + R2 + D1, Supabase, Neon. The right pick for early-stage products that need to ship without a DevOps team.
Infrastructure as code
Terraform, Pulumi, AWS CDK, SST. Reproducible environments, drift detection, secrets management, multi-account and multi-tenant patterns.
Observability & SRE
Datadog, Honeycomb, Grafana, OpenTelemetry, Sentry, Cloudwatch. SLOs, error budgets, on-call runbooks. Production-grade from day one.
Services
Three lines of work, done well.
We're deliberately narrow. Greenfield specialists — building net-new AI systems and platforms from zero is what we do best. Brownfield (existing codebase) work is welcome too, especially Engineering Rescue.
Production AI Applications
LLM-native products, advanced RAG, Graph RAG, agents, and fine-tuned models that survive evaluation.
- Frontier-model integration (Claude, GPT, Gemini)
- Advanced RAG + Graph RAG over your data
- Multi-step agents with guardrails
- Fine-tuning open-weight models
- Eval harness + observability
Bespoke Software Platforms
Custom SaaS, internal tooling, and B2B platforms built for teams that have outgrown templates.
- Typed end-to-end (TypeScript)
- Postgres + your cloud of choice
- Auth, billing, webhooks, integrations
- Documented for your team
Engineering Rescue
Inherited a broken codebase from a previous shop, contractor, or AI-assisted build? We diagnose, stabilize, and ship the plan.
- 10-day technical audit
- Written stabilization plan
- Rebuild-vs-rescue recommendation
- Hands-on remediation sprint
Selected work
Shipped, with metrics.
Most clients prefer not to be named. Industries, architectures, and outcomes are real. References available under NDA on request.
Real-time fraud-triage agent
- Problem
- Manual review queue creating 4-hour decision lag on flagged transactions.
- What we built
- Claude-powered triage agent over Postgres + pgvector with 50k labeled cases, eval harness on every prompt change.
94% auto-resolution
p95 latency 12s · est. $1.4M annual savings · NPS 9.6
HIPAA-aware clinical search
- Problem
- Clinicians spending 40% of charting time searching across 90k unstructured documents.
- What we built
- Hybrid retrieval (BM25 + dense) over partitioned multi-tenant pgvector, citation tracking, on-prem Bedrock deployment for HIPAA.
37% time saved
Sub-second p50 retrieval · zero PHI egress · expanded to 4 hospital systems
Replaced 3 vendor tools with one
- Problem
- Ops team paying $180k/yr for three SaaS tools that didn’t talk to each other.
- What we built
- Custom internal platform on AWS (ECS + RDS + EventBridge), typed end-to-end, integrated with carriers and ERP.
$160k/yr saved
Shipped in 11 weeks · 4× faster ops cycle · zero downtime cutover
Why companies choose us
The difference between shipping and shipped.
Working demo isn't a working system. Here's how we close that gap.
Senior-only delivery
Every line of code shipped by a senior engineer. No juniors hidden in the rate. Our principals do the work, not subordinates.
AI-augmented, not AI-careless
We use AI tooling aggressively to move faster. We do not ship its first draft. That distinction is the entire job.
Fixed-scope, fixed-price
Tiered engagements with a published floor. No hourly drift, no scope-creep games. You know the cost before we start.
You own everything
Code in your repo from day one. Infrastructure in your cloud account. Documentation for your next hire.
What clients say
On the record, off the record.
Quotes anonymized at the client’s request. Names available under NDA after a brief is exchanged.
“We went from "is this even possible" to a working production system in six weeks. The architecture memo alone was worth the engagement fee.”
VP Engineering
Series B B2B SaaS
“Hired Mini Trends after two failed attempts with offshore vendors. They diagnosed our codebase in two weeks, gave us a written rebuild plan, and shipped it in eight.”
CTO
Healthcare startup
“The async model worked exactly as advertised. We never had a standing meeting. Their work product was better than agencies that demanded weekly demos.”
Director of Engineering
Logistics platform
Process
A complete engagement, zero meetings.
Every step happens in writing. You will never be asked to schedule a discovery call, sit through a demo, or attend a standing weekly. The site does the selling. The work product does the rest.
- Phase 01Day 0
You send a written brief
Use the structured form on this site. Ten minutes, no calls. You describe the problem, budget range, and timeline; we do the rest.
- Phase 02Day 1 – 2
We respond in 48 hours, in writing
Either a written proposal — scope, timeline, fixed price, deliverables — or a plain "no, here is why, here is who I would call instead". Always written. Never a sales call.
- Phase 03Day 2 – 5
You sign asynchronously
Standard SOW, e-sign, deposit invoice. No follow-up calls. If you have questions, email them. We answer in writing within one business day.
- Phase 04Day 5+
We ship — async by default
Code lands in your repo from day one. Weekly written updates with demo links. Slack channel for questions. Optional Loom walkthroughs of major milestones. No standing meetings.
Want to see what an async engagement looks like before committing?
Browse the articles for examples of how we write up architecture decisions, eval results, and shipping plans.
Engagement & pricing
Three doors. All priced honestly. All published.
Fixed-scope engagements with a published floor. If your problem doesn't fit a tier, we'll write you back and tell you so.
Most common entry point
Discovery Sprint
$35,000
2 weeks · fixed
A new AI feature, a stalled MVP, or a build-vs-buy question that needs a real answer in writing.
- Architecture audit + written plan
- Working prototype of one critical path
- Risk register & dependency map
- Cost & timeline for the full build
Most engagements land here
Build Engagement
From $75,000
6 – 16 weeks · milestone
Ship a production system. Advanced RAG, an LLM agent, a fine-tuned model, or a full bespoke platform.
- Senior-led delivery, async by default
- Eval harness & observability
- Deployed to your infra; you own it
- Documentation for your next hire
Fractional CTO
Embedded Principal
$30,000 / mo
3-month minimum
Force-multiplier for an existing engineering team. Fractional CTO with hands on the keyboard.
- ~60% hands-on engineering
- Architecture & hiring review
- Async access in your Slack
- Optional board / investor reporting
All engagements are fully async. No discovery calls, no demos, no standing weeklies. Send a brief, get a written proposal in 48 hours.
Procurement & payment
Boutique pricing. Enterprise-ready terms.
Forward this section to procurement. Every common question is answered up front so AP and legal can sign off without back-and-forth.
Fixed-scope, fixed-price
All engagements priced upfront in USD. No hourly drift, no scope-creep change orders. The number on the SOW is the number on the invoice.
Standard payment schedule
50% on signing, 50% on delivery for fixed-scope. Milestone-based invoicing for builds. Monthly in advance for retainers.
Wire / ACH preferred
Wire transfer (USD) or US ACH for engagements over $10k. Credit card accepted under $10k via Stripe. Tax invoices issued for international clients.
Net 15 standard, Net 30 on request
Default invoice terms are Net 15. Net 30 available for enterprise procurement on request — no questions asked.
Vendor docs ready
W-9, W-8BEN, COI ($2M general liability + $1M E&O), MSA template, mutual NDA — provided before contract on request.
Procurement-system friendly
We onboard via Coupa, SAP Ariba, Workday, and similar. Standard vendor packets accepted. Tax-exempt billing handled.
Need something custom? Net 60, milestone variations, tax-exempt billing, escrow, or specific SOW templates — note it in the brief and we’ll accommodate within reason.
Free interactive tools
Use the tools. Skip us if you can.
Free calculators and decision tools we built for our own engagements. No email gate, no signup. If they answer your question on their own, great — that means you didn't need us.
Most popular
AI Cost Calculator
Real 2026 token prices for Claude, GPT-5, and Gemini. See your monthly cost in 30 seconds — and exactly where you’re overpaying.
Open calculator60-second quiz
Build vs. Buy AI
Eight questions, an opinionated written recommendation. Build it, buy a vendor, or go hybrid — with the rationale that fits your situation.
Take the quizFrom the practice
Field notes on production AI.
Long-form essays on what actually ships. The same writing voice we use in proposals — read these to know what to expect.
Production AI
Production AI: What It Actually Takes to Ship LLM Software That Doesn’t Wake You Up at 3 a.m.
A working demo is not a working system. Here is the practitioner playbook for getting LLM applications into production without the usual midnight failure modes.
Read articleEngineering RescueWhen to Rebuild, When to Rescue: A Senior Engineer’s Triage Framework for Failing Codebases
A practical framework for deciding whether to save the codebase you have or start over — and how to make either choice without losing the next two quarters.
Read articleFractional CTOFractional CTO Services: When You Need One, How to Hire One, What to Expect
A founder’s guide to engaging a senior engineering leader part-time — the situations where it works, the situations where it does not, and what to look for in the contract.
Read articleField notes, weekly
One essay a week. No promo.
A single field-note from production AI engineering — one opinionated take, two or three links worth reading. Unsubscribe with a one-word reply.
FAQ
Common questions, answered up front.
The questions procurement, legal, and AP usually ask. If yours is not here, send a brief and ask in the body — we will answer in writing within one business day.
- What does an engagement cost? +
- Discovery Sprint from $35,000 (2 weeks, fixed). Build engagements from $75,000 (6–16 weeks, milestone). Embedded Principal $30,000/month with a 3-month minimum.
- Do you offer discovery calls? +
- No. Our entire engagement is async — no discovery calls, no demos, no standing meetings. You send a written brief, we respond in writing within 48 business hours.
- Who actually does the work? +
- A senior-led practice. The principal answers your brief, runs the engagement, and writes the code. Specialized collaborators are brought in for fine-tuning, security audits, or on-prem deployment phases — never juniors hidden in the rate.
- How is this different from an Anthropic / OpenAI partner shop? +
- Same caliber of work, ~one-third the price ($35k floor vs $150k+), 48-hour written proposal vs 4–8 weeks of sales calls. We are deliberately small and async-first; partner shops are deliberately large and account-team-driven. Both serve real markets.
- What payment terms do you support? +
- Fixed-scope, fixed-price USD. 50% on signing, 50% on delivery for sprints. Wire / ACH preferred over $10k; credit card under $10k. Net 15 standard, Net 30 on request. W-9, COI ($2M GL + $1M E&O), MSA, and NDA available before contract.
- Can I get a refund if it does not work out? +
- Discovery Sprints are first-week refundable. If you do not like the direction by end of week one, we return the deposit, no questions asked.
Why send a brief
Three reasons this is risk-free.
The downside of sending a brief is nothing. The upside is a written, opinionated, senior-engineer take on what you’re building — even if you never engage us.
Free written quote in 48 hours
Send a brief, get a fixed-scope SOW in two business days. Costs nothing. Commits to nothing.
Honest "no" if it’s not us
If your project isn’t the right fit, we say so in writing — and refer you to two or three shops who would be. We turn down work most weeks.
First week refundable
On Discovery Sprints, the first 5 business days are refundable if you don’t like the direction. Full deposit returned, no awkward conversation.
The team
Senior bench. No juniors. No agencies inside the agency.
Mini Trends is a senior-led practice with a small bench of trusted collaborators we bring in for specialized phases — fine-tuning runs, security audits, design sprints, on-prem deployment work. The bench is short and deliberately so. Every engagement has one principal accountable from brief to handover.
Our principal has been writing software professionally since 2004 — through Rails, into React, and now firmly into the age of LLM-native products. Eight years inside production AI before that phrase existed in pitch decks. Past lives include staff roles at venture-backed startups, principal-track consulting for Fortune 500 platform teams, and three of his own companies (one acquired, two graveyard, all instructive).
We are deliberately small and deliberately async. We say no to most engagements because the bar is high and the bench is finite. If you want a fifty-person agency, this is the wrong page. If you want a senior team that delivers in writing and ships production AI without hand-holding — send a brief.
Have a real problem and a real budget?
Send a brief.
Ten minutes of writing on your end. Forty-eight hours of consideration on ours. Either a written proposal or an honest no — never a sales call.
- No discovery calls, ever
- Written proposal in 48 hours
- Fixed scope, fixed price
- You own the code from day one