Rescue and greenfield builds · 4–8 weeks · fixed bid
Rebuild the retrieval system your users stopped trusting.
It demonstrated well. Then real users asked real questions, accuracy landed near 60%, a hallucination was screenshotted, and adoption quietly died. This engagement rebuilds the pipeline for production against a measured accuracy target agreed up front — and leaves your team the harness that keeps it there.
What a rescue rebuilds
- Chunking that respects structure
- Documents split by section and layout; code split by whole functions. Fixed-size splitting is the most common root cause I find.
- Hybrid retrieval and reranking
- Vector and keyword search fused, then reranked for topical relevance to the actual question rather than incidental term overlap.
- Grounding and honesty gates
- Coverage grading, fabricated-citation stripping, and explicit "insufficient information" behavior when retrieval is weak — trust is what returns users.
- Latency and cost engineering
- Semantic caching, escalation gating so expensive steps run only when needed, and model-tier routing. A 41-second worst-case path in my own system now completes in 2.3 seconds.
- An evaluation harness with a CI gate
- A golden set with retrieval-level and answer-level scoring, wired into CI so regressions are caught before users see them.
- A complete handover
- Documentation and a working session so your engineers own the system. The engagement is designed to end.
How the engagement runs
- Week 1Audit phase — or your completed audit carried in: golden set, baseline number, agreed target.
- Weeks 2–4Highest-leverage fixes first, typically chunking and retrieval. Weekly demonstration with the evaluation number moving.
- Weeks 4–6Grounding gates, latency and cost work, and edge cases drawn from your real query log.
- Final weekAcceptance run against the agreed target, CI wiring, and the handover session.
Greenfield builds follow the same shape: the audit phase becomes a design phase and the target is defined on a pilot corpus.
Pricing
$12,000–16,000
Focused rescue · 4 weeks
One pipeline, one corpus type, target accuracy agreed from the audit baseline.
$16,000–25,000+
Full rebuild or greenfield · 6–8 weeks
Multi-source ingestion, agent actions, deployment to your infrastructure, and team training.
from $500 / month
Optional care plan
A monthly evaluation report, drift monitoring, priority fixes, and model upgrades.
Payment is milestone-based, with the final milestone tied to the acceptance run. A RAG Audit completed within the previous 30 days is credited in full.
The advantage you are hiring
Everything above is the working architecture of EAKC, my live production platform of roughly 455,000 lines with over 521 automated tests — where I have personally debugged each of these failure modes and documented the work across 32 recorded engineering episodes. You are not funding my learning curve; you are applying what already works.