AVAI Consulting / Ashvernworks Book a call

Data & search infrastructure

Most AI projects don't fail at the model. They fail at the data.

I build the ingestion pipelines and search systems underneath AI products — for teams whose data arrives messy, from sources they don't control, and whose results have to be right. Systems that run in production, not prototypes that demo well.

Book a 30-minute call

No pitch deck. We look at your actual problem and I tell you whether I can help.

How we can work together

Three ways in

Fixed scope and fixed price wherever possible. If a project doesn't fit one of these, say so on the call and we'll shape it.

Data Pipeline Build

For teams whose data comes from sources they don't own.

Crawlers and ingestion, normalisation, deduplication, staleness detection and monitoring — deployed, documented, and handed over running.

$8,000–15,000FIXED SCOPE · 3–5 WEEKS

Search & Retrieval

For products where search is the product, and the results are wrong.

Hybrid keyword and vector ranking, an evaluation harness so relevance is measured rather than argued about, and tuning against your real queries.

$10,000–18,000FIXED SCOPE · 4–6 WEEKS

Fractional AI Engineering

For teams with a roadmap and nobody to own the hard parts.

Ongoing build and technical direction: architecture decisions, code review, build-versus-buy calls, and the unglamorous work that keeps a system running.

$3,000–5,000PER MONTH · 30 DAYS' NOTICE

Work you can go and look at

Two systems, both live

Not case studies written after the fact — running software you can open in a browser right now.

truepostings.com →

A job aggregation platform built on a persistent company registry rather than live fan-out. Crawls ATS providers, normalises and canonically dedupes postings, detects stale listings, and serves ranked search.

Node · Postgres · Meilisearch · Redis/BullMQ
embeddings · LLM enrichment · live Stripe billing

evchrge.com →

A cost-modelling and incentive database for home EV charger installation. Every published figure is gated behind a verification date — the system structurally cannot assert a number nobody has checked.

Static · zero dependencies · no build step
client-side cost model · verified data layer

Next step

Thirty minutes, and you'll know if this is worth pursuing.

Bring the problem you're actually stuck on. If I'm not the right person, I'll say so on the call rather than scoping something I shouldn't take.

Book a 30-minute call