Sapienbotics

Bid answer libraries

You won it on Friday. They want the evidence by Wednesday.

Your approved answers already exist — in last year’s submission, in a folder nobody can search, in the head of whoever wrote them. Rebuilding them is a search-and-adapt job that eats evenings, and it happens again on the next bid. I build the library that holds each answer with the evidence that proves it, so the next application starts from what you already own. It will not win you a contract. It gives you back the evenings.

Saurabh Singh · MTech and PhD, IIT Delhi · systems for bid teams · United Kingdom

Built for: Bid and Commercial Managers at 20–500 person firms
Sectors: Construction, engineering, facilities, professional services
Delivery: Four weeks, fixed price

01Your last bidsSubmissions, applications
02ExtractOne record per answer
03LibraryEvidence + expiry date
04Evidence gateUnevidenced content flagged
05AssembleCurrent content only
06Your reviewAlways a person
→First draftCurrent, and sourced

How the system runs, end to end. Your last bids → Extract → Library → Evidence gate → Assemble → Your review → First draft

The offer

A bid answer library. Four weeks. £4,500 fixed, once.

One system, one price, stated openly because you should not have to book a call to find out whether you can afford it. You own what I build and it runs on your infrastructure. There is no seat count, no renewal, and nothing to cancel.

Price: £4,500
Terms: One time, no renewal
Duration: 4 weeks
Runs on: Your infrastructure

What it does

Extracts every reusable answer from your last submissions and framework applications — accreditations, insurances, policies, method statements, personnel, project experience, financial and health-and-safety records.
Holds each one with the evidence file that proves it and the date it stops being true.
Marks any answer it cannot evidence rather than writing a sentence that reads well and is not supportable — a NEEDS EVIDENCE flag in the draft, for you to resolve or drop.
Assembles a structured first draft for a framework application or ITT from current content only, with every answer traceable to the library item it came from.
Keeps the post-award evidence pack assembled and current, so the request that follows a win is a lookup rather than a scramble.

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, and it does not quietly substitute something close. It stops and names the gap.
It does not judge whether a claim is good. It establishes whether it is current and evidenced. Those are different jobs and only the second one can be automated honestly.
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.

What it costs, in the unit you already use. Outsourcing one public tender to a bid writer costs roughly what this does — the bid-writing trade puts the average at about £4,500 a bid, and a specialist day rate of £600 buys you seven or eight days. That buys one submission. Then the next ITT arrives and you buy it again.

This is that same fee, spent once, on the asset instead of the event. You get no written bid out of it — say that plainly — and it will not win you a contract on its own. What you get is that the second bid, and the ninth, start from answers you already own and can evidence, rather than from a search through last year’s submissions. If you bid twice a year, buy the bid writer. If you bid often enough that the same answers get rebuilt from scratch, buy the library.

Selected work

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

Read what follows as 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. Everything below is what I can evidence and nothing more; judge the rest on a call.

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 ranked without a developer in the loop: ~2,200companies ranked without a developer in the loop
needed to change the model — the team edits it in conversation: No devneeded to change the model — the team edits it in conversation
batch refresh cycle: 12 minbatch refresh cycle

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.

Tender and market intelligence

Indian infrastructure contractor

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

searched by hand every morning, now one bulletin at 08:00: 5 sourcessearched by hand every morning, now one bulletin at 08:00
records ingested → the 59 worth a person reading: 582 → 59records ingested → the 59 worth a person reading
stage failures across all seven stages at handover: Zerostage failures across all seven stages at handover

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.

Email and order automation

European construction-supply 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.

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.

scanned survey plats, before — now computed and checked: Read by handscanned survey plats, before — now computed and checked
of acreage figures gated for human review, by design: 100%of acreage figures gated for human review, by design
phases delivered in sequence: 3phases delivered in sequence

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.

The wider portfolio

Twenty-two delivered systems, and what each one had to get right.

One thing is for sale on this page and it is the bid answer library above. This list is here for a different reason: so you can check whether I have built anything shaped like your problem before you spend a call finding out. Grouped by the shape of the problem rather than by industry, because that is the axis you are actually scanning.

No client is named. One prior engagement is governed by an agreement I have read and one document I have not, so every description here is anonymised and one identifying detail is blurred in each case where the sector alone would give the client away. I will say more on a call than I will publish.

Documents in, structured data out

05The pattern the bid library belongs to.

01Public-sector procurementGovernment RFP response enginea US public-safety staffing companyIngests a municipal RFP with its addenda, retrieves approved content, and renders a submission-ready PDF and DOCX.
Placeholder names hard-fail the render, and the generating agent is forbidden from certifying its own output.
Python on Cloud Run · n8n · Next.js · Supabase · OpenAI · LibreOffice headless
02Construction supplyEmail and order automationa European construction-supply distributorReads inbound mail and PDF attachments, extracts the order data, and routes it across two CRMs and Slack.
Several hundred emails a day, full EU data residency, and three plausible features named out of scope in the SOW rather than absorbed.
Self-hosted n8n · Supabase (EU) · GPT-4.1 vision in strict JSON · Gmail API · HubSpot and a second CRM
03Oil and gasMineral-rights acquisition pipelinea Texas mineral-rights buyerReads scanned land-survey plats to compute net royalty acres, then generates the purchase contract.
Bearings read off scans and section geometry inferred proportionally — with every computed figure gated for human sign-off before it reaches a legal instrument.
n8n orchestrator and four sub-workflows · GPT-4o vision · Drive API · DOCX merge · LibreOffice
04Luxury travelInternal knowledge assistanta UK luxury travel operatorIngests four source systems nightly and answers staff questions with citations back to the source.
Deploys gated on four Puppeteer suites, and a pre-handover scan for PII, credentials and passport numbers before anything left my machine.
Next.js · Supabase · n8n · CRM, itinerary-platform, mail and Drive connectors · Puppeteer
05Security complianceCompliance account discoverya US software companyRead-only factual audit of a compliance-platform account, run against a beta API before a proposal was finalised.
Deliberately scoped as discovery: no writes, no architecture, findings only. It is on this list because knowing when not to build is part of the record.
GraphQL beta API · schema introspection

Generate the client’s document and deliver it

04Assembly and delivery, not drafting assistance.

06Agency reportingMarketing reporting platformown product, ceasedAgencies connect ad and analytics accounts; the system writes the commentary and emails a branded deck on a schedule.
Shipped complete and then ceased: it depended on one platform’s ad integration and that was rejected at app review. Permissionless distribution has been a hard constraint on everything since.
Next.js · FastAPI on Railway · Supabase with RLS · GPT-4.1 · python-pptx · matplotlib · AES-256-GCM
07Market researchSurvey methodology agenta B2B services companyDrafts qualitative survey outlines and self-audits them against a written methodology rulebook before export.
The rulebook is the actual deliverable. The model revises against it and the audit runs before anything is exported.
n8n · Claude for generation and revision · Node docx · Google Docs API · Firebase
08EducationBusiness-plan assessment agentWalks a student through a five-section business-plan framework, scores it against rubrics, and emails the compiled report.
Rubric scoring rather than judgement: every score traces back to the section and criterion that produced it.
n8n · Voiceflow · GPT-4 · Supabase · Gmail API
09EducationEducational video generationown productTurns a topic prompt into a narrated educational video — scene plan, dual-engine render, assembly, narration.
97 parametric Manim generators across eleven subject domains, alongside 78 React animations.
Node and Express · React with Puppeteer · Manim · FFmpeg · Three.js · GPT-4o · Cloud Run

Score, rank and route

04Where a number decides what happens next.

10Venture capitalDeal prioritisation systema US venture capital firmScores CRM companies across five weighted dimensions and lets the investment team change the model in conversation.
Around 2,200 companies on a twelve-minute cron. The passcode gate is enforced in the database through a session GUC and RLS, so the model cannot talk its way past it.
n8n Cloud · Supabase Postgres, 13+ migrations with RLS · CRM and market-data APIs · GPT-4.1-mini · MCP
11Education staffingTutor matching and outreachMatches tutors to students and runs the messaging campaign that follows.
The system never ranks or selects. It surfaces candidates and a person decides — which was the client’s requirement, and the right one.
Next.js · NestJS · Prisma · Azure Postgres · Entra ID · Azure OpenAI · WhatsApp Business · SharePoint Graph
12Retail fintechAlgorithmic trading systemown productAutomated swing trading on listed equities, with the model as gatekeeper rather than predictor.
Version 1.0 annotated itself “dangerous — 1.5× Kelly”. Corrected to Kelly 1.0× with decay to zero after four losses, regime switching and a volatility kill switch.
Cloud Functions Gen 2 · Firestore · n8n · broker API · FlutterFlow dashboard · Telegram alerts
13B2B demand generationOutbound acquisition infrastructurea practitioner-coaching businessCold-email acquisition plus a daily content engine across five buyer personas.
Fully engineered and never activated — zero sends. Its teardown is why this practice now sends ten by hand before it builds anything at all.
n8n, ten workflows on the client’s VPS · Smartlead · GoHighLevel · Claude and OpenAI · Apify · Clay

Reach a system that has no usable API

02The work nobody scopes and everybody needs.

14Construction and infrastructureBusiness-development intelligence bulletinan Indian infrastructure contractorIngests tenders, exchange announcements and news daily, then delivers a classified bulletin by messaging and email.
A portal with no API and an OTP login: the browser signs in, the portal emails a code, a mail rule forwards it, the code lands in a table and the scraper polls inside a two-minute window.
Node and TypeScript on a VPS · Supabase with pgvector · OpenAI · Playwright and Cheerio · Next.js 14
15Real estate and mortgageCRM consolidation and voice qualificationa US real-estate and mortgage operatorConsolidated one contact base spanning three businesses and put a voice agent on inbound qualification.
OCR run against email threads because the source system had no export worth trusting. Unglamorous and the only thing that worked.
CRM API · voice-agent platform · CSV reconciliation · OCR

Agents that act, and cannot exceed a limit

03The constraint is the product.

16Local commerceLocal commerce marketplaceown productA buyer describes what they want, the system matches a nearby business, and an agent negotiates the price.
The agent is bounded by per-product commercial rules — minimum price, maximum discount, minimum quantity. It can negotiate and it cannot concede past the owner’s floor.
FlutterFlow, 18+ custom widgets · n8n, a 65-node workflow · Groq Llama 4 and Whisper · SerpAPI · Firebase
17Voice AI and sales operationsVoice agent and post-call chainTakes the call and runs the automation that follows it, built as a reusable reference implementation.
Built to be lifted into later engagements rather than delivered once and abandoned.
Retell · n8n · Supabase · HubSpot · GoHighLevel · 1Password
18Research protocolProduct validation protocolown methodA six-phase protocol that mines public sources for user pain and kills candidates down to one validated idea before any code is written.
535 raw signals to 21 clusters to eight surviving competitor checks to three finalists. Synthetic data banned outright, with a stop-gate at every phase.
Claude Code as execution engine · Python analysis · public-source collection

Applications and interfaces

04Where somebody has to actually use it.

19Language learningPronunciation coaching appown productListens to spoken English, scores the attempt, and a tutor agent responds with a correction in two languages.
The score is signal processing, not an API call: dynamic time warping over MFCC features, per-user calibration from a warm-up, and thresholds that move with CEFR level.
Flutter, 30+ custom widgets · Python DTW service on Cloud Run · MFCC, CMVN, VAD · Groq · Google TTS
20Consumer AIAI creative tools appown productSocial chat alongside five creative tools on a credit subscription.
Three GPU models orchestrated async submit-and-poll inside a stateless workflow engine, with atomic credit deduction per generation.
FlutterFlow · n8n, seven workflows · RunPod Serverless (FLUX, Wan) · GPT-4o-mini · Firebase, 14 collections
21HospitalityGuest concierge platforma hospitality operatorPhased guest-facing concierge with a super-admin dashboard, guest sign-up and multi-language support.
The voice reservation agent was scoped out and deferred rather than promised into the first phase.
FlutterFlow, four registered widgets · n8n, six workflows · Python patch scripts
22Practice operationsPractice management systemown toolingConversational project management for this practice against Sheets and Drive — a schema and an operating manual, not an application.
Encodes a payment lifecycle with an explicit rule never to auto-advance on elapsed time, and a read-only boundary over every other project folder.
Claude Code with MCP over Sheets and Drive · Google APIs · Python

22 projects, twenty-one of them delivered inside 2026. That is velocity and breadth, not tenure, and it is the honest way to describe it.

How I work

Find the step that cannot be trusted, and put a person in front of it.

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.

On one proposal system I built, an appendix was inventing personnel names — in seventeen of twenty proposal families. The fix that matters is the last one: a check that hard-fails the render if an invented name ever reaches the output. The whole corpus was then re-rendered and independently re-audited to zero. A warning someone has to notice is not a control; a failed build is.

That is the same mechanism, pointed at a different failure. A bid library’s version of an invented name is a case study naming a site manager who left, a policy superseded last year, or a figure nobody can now source. None of those look wrong on the page. So the library holds the evidence against every claim, and the assembly step flags what it cannot stand behind rather than proceeding quietly. That is a build standard, not a sales pitch: it is there because I will not ship a system that writes confident sentences it cannot support.

Next step

Tell me what you are bidding next.

I will tell you plainly whether a library helps — and if you bid twice a year, it does not, and I will say so and that is the end of it. Paste the framework application or the tender. If it sits behind a portal, the reference and the closing date are enough to start.

Email me hello@sapienbotics.com

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