Sapienbotics

Bid-response systems

Your bid team keeps rewriting answers it has already written.

Every ITT asks the same twenty questions in a different order. The approved answers already exist — in last year’s submission, in a folder nobody can search, in the head of whoever wrote them. I build the system that finds them and assembles the first draft.

Built for
Heads of Bids at 20–500 person firms
Sectors
Construction, engineering, professional services
Delivery
Four weeks, fixed price
How a bid-response system processes a tenderITT / RFP → Split → Match → Assemble → Your review → First draftITT / RFPPDF, Word, portal export01SplitOne record per question02MatchYour approved library03AssembleStructured first draft04Your reviewAlways a person05First draftSourced, not generated
How the system runs, end to end. ITT / RFP → Split → Match → Assemble → Your review → First draft

The offer

A bid-response system. Four weeks. £4,500 fixed.

One system, one price, stated openly because you should not have to book a call to find out whether you can afford it.

Price
£4,500
Currency
GBP, fixed
Duration
4 weeks
Runs on
Your infrastructure

What it does

  • Ingests the ITT, RFP or tender pack and splits it into individual questions and requirements.
  • Matches each question against your own approved content library — your words, previously signed off.
  • Assembles a structured first draft with every answer traceable to the source it came from.
  • Runs deterministic compliance checks: word limits, mandatory sections, missing responses.
  • Flags what it could not confidently answer instead of filling the gap with something plausible.

What it does not do

  • It does not write your bid. It assembles a first draft from answers you have already approved.
  • It does not invent content when the library has no match. It tells you the gap exists.
  • It does not submit anything. Nothing leaves your hands without you reading it.
  • It does not need a new CRM, a new portal, or a change to how your team works.

Selected work

Four systems, and the safeguard that made each one trustworthy.

01

Deal prioritisation

US venture capital firm

An investment team ranking its own pipeline by hand, with no way to change the ranking model without a developer.

companies scored, tiered and explained
~2,200companies scored, tiered and explained
batch refresh cycle
12 minbatch refresh cycle
weighted scoring dimensions
5weighted scoring dimensions

The safeguard

Changing the scoring weights rewrites the firm’s view of its own pipeline, so it is passcode-gated in the database rather than in the prompt — a SECURITY DEFINER function sets a session variable and row-level security checks it. There is no instruction that talks the model past it, because the model is not what holds the gate. Every change is versioned and every rollback saves the pre-rollback state first.

n8n Cloud · Postgres · CRM API · GPT-4.1-miniHandover package the client deploys and runs
02

Tender and market intelligence

Indian infrastructure contractor

A business development team searching five sources by hand every morning, two of them public procurement portals.

records ingested → passed the relevance gate
582 → 59records ingested → passed the relevance gate
stage failures across all seven stages at handover
Zerostage failures across all seven stages at handover
daily bulletin, WhatsApp and email
08:00daily bulletin, WhatsApp and email

The safeguard

The 523 records that did not pass the relevance gate are retained and flagged, not deleted — so the team can ask why something was excluded and get an answer, and a filter that is too aggressive is visible instead of invisible. Low-confidence classifications go to a human review queue rather than into a discard, and a pipeline failure raises an alert rather than producing a quiet morning with no bulletin.

Node/TypeScript · Postgres + pgvector · Playwright · Next.jsRunning unattended on the client’s own server
03

Email and order automation

Dutch building-products distributor

300 to 500 emails a day processed by hand, across two CRMs, a supplier portal, and inbound PDF drawings and invoices.

emails a day previously handled manually
300–500emails a day previously handled manually
CRMs kept correct as one order object
2CRMs kept correct as one order object
server, database and storage, by design
EU-onlyserver, database and storage, by design

The safeguard

Three plausible, adjacent capabilities are named in the contract as explicitly out of scope: no browser automation against the supplier portal, no drawing generator, no custom dashboard. On a fixed-fee project, writing down what is not being built is what keeps the thing deliverable. Classification and PDF parsing both run in strict JSON mode, so a malformed response is a caught error rather than a bad record written into a CRM.

Self-hosted n8n · Postgres · object storage · Gmail, two CRMs, SlackSelf-hosted on the client’s own EU server
04

Land document analysis

Texas mineral-rights buyer

Scanned land survey plats read by hand to compute royalty acreage, then purchase contracts drafted from a spreadsheet.

of computed acreage figures gated for human review
100%of computed acreage figures gated for human review
phases: plat reader, pipeline, operating view
3phases: plat reader, pipeline, operating view
sub-workflows under one orchestrator
4sub-workflows under one orchestrator

The safeguard

A scanned survey gives bearings and distances, but the section geometry often has to be inferred proportionally, and the document it flows into becomes a legal instrument with a royalty calculation in it. So the system computes, states its basis, and stops for a person. On a legally binding document an honest confidence signal is worth more than a point of accuracy: a system right 97% of the time and silent about which 3% is worse than one right 94% of the time that tells you where it was unsure.

n8n · GPT-4o vision · LibreOffice headless · CRM outreachThree phases, delivered in sequence

How I work

Twenty-five systems in six months, and one thing I will not do.

I build working systems on your infrastructure and hand them over with the documentation to run them without me. Fixed scope, fixed price, and what is *not* being built written down before it starts — on a four-week engagement that is the discipline that keeps it deliverable.

The pattern in the four projects above is the same: find the step that cannot be trusted, isolate it, and gate it behind a person. The plat reader computes acreage and then stops. The scoring model ranks and explains but does not decide. The relevance filter keeps what it rejected so you can audit it. Automation that hides its own uncertainty is worse than no automation, because you stop checking.

One thing about the numbers above: they are delivery metrics — records processed, cycles run, stages passing at handover. They are not business outcomes. I do not have verified figures for tenders won, hours saved or revenue moved, because those were measured after handover by the client and were not reported back to me. I could estimate them. Every agency site does. I would rather tell you exactly what I can evidence and let you judge the rest on a call.

Next step

Send me a tender you are working on.

I will tell you plainly whether a system helps — and if the answer is no, I will say so and that is the end of it. Paste the document or a link. If it sits behind a portal, the ITT reference and the closing date are enough to start.

No pitch deck, no discovery sequence, no follow-up cadence.